Towards Evaluation Engineering: An Empirical Study of ML Evaluation Harnesses in the Wild
Paper • 2605.24213 • Published • 3
Datasets:
Formats:
json
Languages:
English
Size:
10K - 100K
ArXiv:
Tags:
software-engineering
evaluation-harness
github-issues
issue-classification
empirical-software-engineering
llm-evaluation
License:
harness_name stringclasses 56
values | github_repo stringclasses 59
values | issue_title stringlengths 1 728 | issue_body stringlengths 0 65k ⌀ | issue_url stringlengths 39 77 | issue_created_at int64 1,376B 1,768B | issue_closed_at int64 1,382B 1,776B ⌀ | is_related float64 0 1 ⌀ | stage float64 0 4 ⌀ | step stringclasses 3
values | strategy float64 1 6 ⌀ | root_cause_label stringclasses 11
values | issue_comments listlengths 0 76 | issue_labels listlengths 0 9 ⌀ | issue_cross_referenced listlengths 0 23 ⌀ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
ranx | AmenRa/ranx | Issue with MRR | Attempting to the MRR with your example and am getting a Typing error.
System:
- python 3.8.2
```
from rank_eval import ndcg, mrr
import numpy as np
# Note that y_true does not need to be ordered
# Integers are documents IDs, while floats are the true relevance scores
y_true = np.array([[[12, 0.5], [25, 0... | https://github.com/AmenRa/ranx/issues/1 | 1,620,665,993,000 | 1,621,426,763,000 | 1 | 3 | A | 1 | Validation Gap | [
"It works on my end.\r\n\r\nDo the other metrics raise the same error?",
"Closing for inactivity. Feel free to reopen the issue."
] | null | null |
ranx | AmenRa/ranx | bug: using built-in `type` function instead of type argument | It’s all in the title :)
https://github.com/AmenRa/ranx/blob/f41e64d92e62a6ecb484f1271b481ec581aa05f9/ranx/run.py#L140 | https://github.com/AmenRa/ranx/issues/10 | 1,643,126,681,000 | 1,643,127,231,000 | 1 | 3 | A | 1 | Validation Gap | [
"Thank you. It should be fixed now (0.1.8). Closing."
] | null | null |
ranx | AmenRa/ranx | set Report’s rounding_digits in compare | Hi,
`compare` does not have a `rounding_digits` argument and thus always uses the default from `Report` (which is 4). Why is that?
Also, would you like to add an option in Report to print results as percentages rather than ratios ? | https://github.com/AmenRa/ranx/issues/11 | 1,644,252,698,000 | 1,644,946,451,000 | 1 | 4 | A | 1 | Unimplemented Feature Gap | [
"Hi Paul,\r\n\r\nI have chosen the number of digits based on the max number of digits I found in IR / SecSys papers for reporting performance scores. However, I was planning on adding an argument as you suggested, but it seems I forgot about it. My bad.\r\n\r\nRegarding using percentages instead of ratios, do you m... | null | null |
ranx | AmenRa/ranx | [Feature Request] Benchmarking | **Is your feature request related to a problem? Please describe.**
To claim "blazing-fast", it would be nice to have benchmarks against existing implementations.
**Describe the solution you'd like**
The implementation is benchmarked against some/all of the sources below:
1. naive [numpy implementation](https://gi... | https://github.com/AmenRa/ranx/issues/12 | 1,645,020,560,000 | 1,656,079,434,000 | 1 | 3 | A | 4 | Unimplemented Feature Gap | [
"Hi, thank you for your interest in `ranx`!\r\n\r\nI compared `ranx` against `pytrec_eval` for my ECIR paper, you can re-run the benchmark using this [notebook](https://github.com/AmenRa/ranx/blob/master/dev/efficiency_test.ipynb).\r\n\r\nHere are the results reported in the ECIR paper that will be out in April:\r\... | null | null |
ranx | AmenRa/ranx | [Feature Request] Reciprocal Rank Fusion | Hi,
Thank you for this nice library.
How would you go about implementing [Reciprocal Rank Fusion](https://plg.uwaterloo.ca/~gvcormac/cormack) in ranx?
Thanks,
Maxime. | https://github.com/AmenRa/ranx/issues/13 | 1,648,133,901,000 | 1,648,639,919,000 | null | null | null | null | null | [
"Hi Maxime and thanks for your question,\r\n\r\nI plan to add many fusion methods to `ranx` this year, but I am currently busy with other projects.\r\n\r\nRight now, your best bet is to implement `RRF` with Python dictionaries and pass the resulting fused run to `ranx `.\r\n\r\nAlternatively, you can use an availab... | [
"enhancement"
] | [] |
ranx | AmenRa/ranx | [Feature Request] from ranx import Report | **Is your feature request related to a problem? Please describe.**
Hi! I’d like to be able to import Report so that I can load a previously saved report (output of `compare`) and tweak the runs.
**Describe the solution you'd like**
[`from .report import Report`](https://github.com/AmenRa/ranx/blob/master/ranx/__in... | https://github.com/AmenRa/ranx/issues/16 | 1,651,654,127,000 | 1,651,754,301,000 | 1 | 4 | A | 5 | Unimplemented Feature Gap | [
"mmh actually it’s not straightforward to load the saved report into a new one. it would need something like a `from_dict` method that reverses `to_dict` (https://github.com/AmenRa/ranx/blob/master/ranx/report.py#L240) ",
"Hi Paul,\r\n\r\nPlease, give me a more detailed description of the use case.\r\nWhat do you... | null | null |
ranx | AmenRa/ranx | [BUG] win_tie_loss in Report.to_dict | **Describe the bug**
When converting a Report to dict, you only keep one `metric` while iterating over `metrics` (overwriting the previous `metric` in each loop)
https://github.com/AmenRa/ranx/blob/master/ranx/report.py#L315
**How to fix**
replace the line above with:
`d[m1]["win_tie_loss"][m2][metric] = self.... | https://github.com/AmenRa/ranx/issues/17 | 1,652,965,102,000 | 1,652,966,212,000 | 1 | 4 | A | 5 | Algorithmic Error | [
"Well spotted, thanks!\r\n\r\nFixed in `v0.1.14`."
] | null | null |
ranx | AmenRa/ranx | [BUG] from_ir_datasets seems to be missing from Qrels | **Describe the bug**
Qrels.from_ir_datasets("msmarco-document/dev") does not seem to work, it seems to have been removed by a commit 10 days ago. (https://github.com/AmenRa/ranx/commit/e0ca82c07392d90998b298515f48d0059c48c93d#)
| https://github.com/AmenRa/ranx/issues/18 | 1,655,396,055,000 | 1,655,397,029,000 | 1 | 1 | C | 1 | Unimplemented Feature Gap | [
"Hi and thank you very much for pointing it out.\r\nI must have used an old version of the code when I did the refactoring from the commit you mentioned, gotta check everything... :S\r\n\r\nI restored \"from_ir_datasets\" in `v0.2.4`.\r\n\r\nThanks again!",
"Thanks a lot for this library, just came back after a w... | null | null |
ranx | AmenRa/ranx | Why in Recall@k you divide on len(relevant), but not min(len(relevant), k) | The question about Recall@k arose when I looked at the best scores R@1 of Stanford Online Products dataset in paperswithcode https://paperswithcode.com/sota/metric-learning-on-stanford-online-products-1. This benchmark use R@1 metric to measure best models and approach in retrieval task in SOP dataset. Sop dataset has ... | https://github.com/AmenRa/ranx/issues/19 | 1,656,482,593,000 | 1,656,794,495,000 | 1 | 3 | A | 1 | Documentation Deficiency | [
"Hi, that's an interesting question!\r\n\r\n`ranx` is built to reproduce [`trec_eval`](https://github.com/usnistgov/trec_eval)'s scores, as it is the standard evaluation library used in Information Retrieval research.\r\n`ranx`'s `recall@k` works as [`trec_eval`](https://github.com/usnistgov/trec_eval)'s `recall@k`... | null | null |
ranx | AmenRa/ranx | Problem with r-precision | Hi,
I tested your code and found that it was easy to use and integrate. Moreover, the results I got are fully coherent with those I previously obtained with a personal implementation of trec_eval and the computation of the measures is fast. This is clearly an interesting software and its presentation to the demo ses... | https://github.com/AmenRa/ranx/issues/2 | 1,638,262,771,000 | 1,638,382,547,000 | 1 | 3 | A | 1 | Validation Gap | [
"Hi Oliver,\r\n\r\nThanks for your interest in `rank_eval` and the kind words.\r\n\r\nYesterday, when I did the last commit, I noticed something was off in the code there!\r\nI'm going to address the problem in the next few days and come back to you.\r\n\r\nThanks for your feedback.\r\n\r\nHave a good one,\r\n\r\nE... | null | null |
ranx | AmenRa/ranx | Spearman, Kendall correlation functions | Hi! This lib is extremely good tool to have in arsenal, but I think it could be nice to have Spearman and Kendall correlation functions included to this package to evaluate ranking. Maybe not the most popular metrics, but sometimes they could come in handy.
Best regards,
Ivan Savchuk | https://github.com/AmenRa/ranx/issues/20 | 1,658,404,191,000 | null | 1 | 3 | A | 1 | Unimplemented Feature Gap | [
"Hi and thanks for the kind words! :D\r\n\r\nKendall, Spearman, Pearson, and AP correlation metrics are already in my local branch and are coming soon!\r\n\r\nIf you haven't, please consider giving `ranx` a star. \r\n\r\nBest,\r\n\r\nElias",
"Any update on this?",
"Hi, sorry for the delay.\r\nI will work on thi... | null | null |
ranx | AmenRa/ranx | Issues with MAP | null | https://github.com/AmenRa/ranx/issues/21 | 1,659,384,060,000 | 1,659,384,122,000 | 1 | 3 | A | 1 | Others | [] | null | null |
ranx | AmenRa/ranx | Problems with MAP | I understood that, when evaluating MAP@k, relevance judgment scores equal to 0 are ignored.
In my case, I get a bit of a weird behaviour.
I'm working on a balanced dataset with binary relevancy and define qrels by including both 1s and 0s documents.
While ndcg@10 gives me results at about 0.7, MAP@10 is extremely ... | https://github.com/AmenRa/ranx/issues/22 | 1,659,384,145,000 | 1,661,764,404,000 | 1 | 3 | A | 1 | Documentation Deficiency | [
"Hi Stefano,\r\n\r\nAlmost all the metrics ignore qrels with zero scores, including NDCG.\r\nSo the difference you get is not because of that.\r\n\r\nHowever, I think your results are entirely possible if you have many relevance judgments for each query.\r\nNote that the [Average Precision](https://amenra.github.io... | null | null |
ranx | AmenRa/ranx | Why ranx is too slow in this simple example? | ```python
from ranx import Qrels
from ranx import Run
from ranx import evaluate
qrels_dict = {
"text_1": {
"label_1": 1
},
"text_2":{
"label_2": 1,
}
}
qrels = Qrels(qrels_dict, name="testing")
run_dict = {
"text_1": {
"label_1": 1,
"la... | https://github.com/AmenRa/ranx/issues/23 | 1,659,562,753,000 | 1,659,636,829,000 | 1 | 3 | A | 1 | Documentation Deficiency | [
"Wow, it is taking forever in a real example with about 60k queries!\r\nAm I missing something?",
"Hi, and thanks for your interest in `ranx`!\r\n\r\nI think you are missing that all [`numba`](https://numba.pydata.org)-based `ranx` functions need to be compiled the first time you use them (_there is a disclaimer ... | null | null |
ranx | AmenRa/ranx | [Feature Request] infer run/qrels format from file extension in `from_file` | **Is your feature request related to a problem? Please describe.**
I think it’s quite frustrating to have to specify the format of qrels/run (TREC or JSON). I often get exceptions if I forget to specify the 'trec' format because JSON is default.
**Describe the solution you'd like**
You could infer the format from ... | https://github.com/AmenRa/ranx/issues/24 | 1,660,302,528,000 | 1,661,770,516,000 | 1 | 1 | B | 1 | Unimplemented Feature Gap | [
"Hi Paul,\r\n\r\nI agree with you, and I thought about it in the past.\r\n\r\nHowever, I plan to have a custom JSON-based file for `Run`s (_in addition to those already supported_) and using the sole file extension would not be enough/safe to infer the input file type.\r\nSo, do not bother opening a PR, as your cod... | null | null |
ranx | AmenRa/ranx | [BUG] ttest_rel() got an unexpected keyword argument 'alternative' when using compare with stat_test="student" | **Describe the bug**
Hi, I’m having an error when using `compare` with `stat_test="student"` (no problem when using the default "fisher").
```py
TypeError Traceback (most recent call last)
<ipython-input-3-0369b81922de> in <module>
7 metrics=["map@100", "mrr@100", "ndcg@... | https://github.com/AmenRa/ranx/issues/25 | 1,661,533,569,000 | 1,661,767,603,000 | 1 | 3 | B | 1 | Environment Incompatibility | [
"Hi, I suspect your `SciPy` version is lower than `1.6.0` (see [here](https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.ttest_rel.html)).\r\n\r\nI will explicit the required `SciPy` version in `ranx`'s `setup` file so that this will not happen again.\r\n\r\nIn the meantime, updating your `SciPy` inst... | null | null |
ranx | AmenRa/ranx | Qrels and Run query ids do not match | **Describe the bug**
I am evaluating my test set against my algorithm recommendations, but it gives the error
Qrels and Run query ids do not match
**To Reproduce**
Steps to reproduce the behavior:
metrics = ["recall@10", "mrr@10","ndcg@10"]
person_date_indexs = df_recs_train_top10['person_date_index'].unique()... | https://github.com/AmenRa/ranx/issues/27 | 1,663,436,358,000 | 1,664,264,221,000 | 1 | 3 | A | 1 | Validation Gap | [
"Does your run contains all the elements on the qrel (and vice-versa)? I had this problem many times working with a dataset that had many queries that did not have relevance labels. It can also happen in the opposite sense (although rarer), if you have relevance data for a query that is not evaluated by your system... | null | null |
ranx | AmenRa/ranx | Ranx pip installing failed | **Describe the bug**
Error during pip install.
**To Reproduce**
`pip install ranx==0.3.2`
**Bash output**
```python
(xCoFormer) celso@capri:~/projects/xCoFormer$ pip install ranx==0.3.2
Collecting ranx==0.3.2
Using cached ranx-0.3.2-py3-none-any.whl (93 kB)
Requirement already satisfied: numpy in /home/c... | https://github.com/AmenRa/ranx/issues/28 | 1,666,703,311,000 | 1,667,574,294,000 | 1 | 0 | A | 2 | Environment Incompatibility | [
"Hi,\r\n\r\nI tried to reproduce the error, but the installation works on my end.\r\n\r\nBy reading your reported pip output, the error arises when installing [pyautocorpus](https://github.com/seanmacavaney/pyautocorpus).\r\n\r\nI suggest you to retry installing ranx. If it fails, try to install pyautocorpus.\r\nIf... | null | null |
ranx | AmenRa/ranx | [Feature Request] relevance_level parameter | I was wondering if similar to trec_eval that we can specify the relevance_level, with -l parameter, this feature exists in ranx. If not that would be a useful feature for evaluation | https://github.com/AmenRa/ranx/issues/29 | 1,667,172,328,000 | 1,667,401,189,000 | 1 | 3 | A | 1 | Unimplemented Feature Gap | [
"Dear Behrooz,\r\n\r\nThanks again for your interest in `ranx`! :)\r\n\r\nI added the requested feature in `v0.3.3`.\r\nI added two different ways to set relevance levels.\r\n\r\n__Qrels-wise__\r\nQrels now provides the `set_relevance_level` function: `qrels.set_relevance_level(rel_lvl)`.\r\nIt changes the document... | null | null |
ranx | AmenRa/ranx | confusing variable | https://github.com/AmenRa/rank_eval/blob/master/rank_eval/meta_functions.py#L220
`control_metric_scores` could be defined outside the `for j` loop :) | https://github.com/AmenRa/ranx/issues/3 | 1,639,585,728,000 | 1,640,012,345,000 | null | null | null | null | null | [
"Well spotted! Thank you for taking the time to review the code!\r\nI will fix it soon.",
"Fixed. Closing."
] | [] | [] |
ranx | AmenRa/ranx | [BUG] `MRR@1` is not equal `Recall@1` | **Describe the bug**
`MRR@1` should be equal to `Recall@1`. However, these metrics diverge for the case below.
**To Reproduce**
```python
%%capture
!pip install ranx
from ranx import Qrels, Run, evaluate
import pickle
# download files from https://drive.google.com/drive/folders/1ZLyB6mKKiQsypw36nhdZ4dGqmF... | https://github.com/AmenRa/ranx/issues/30 | 1,667,567,334,000 | 1,667,571,934,000 | 1 | 3 | A | 1 | Documentation Deficiency | [
"MRR@k is only equal to Recall@k if you have exactly one positive per query. Unfortunately, there are many meanings for Recall, but at least for trec eval (which ranx follows) it always uses positives_found/total_positives, even if k is smaller than the amount of positives you have. \r\n\r\nFor the other definition... | null | null |
ranx | AmenRa/ranx | [BUG] Precision calculation incorrect? | **Describe the bug**
In the below example, I would expect run1 to have a precision of 1.0 and I would expect both run2 and run3 to have precisions of 0.75, as 3 out of 4 returned documents are relevant. Instead the second query returns 0.5, and the third 0.25. Either there is a bug handling empty query results, or I h... | https://github.com/AmenRa/ranx/issues/31 | 1,669,067,501,000 | 1,669,111,309,000 | 1 | 3 | A | 1 | Algorithmic Error | [
"It seems that if there is an empty query result in the run_dict, every query after it will always have a precision of 0.",
"Hi @kaleko,\r\n\r\nThank you very much for the bug report and for providing a working example!\r\n`numba` was not raising a `ZeroDivisionError`, so I did not spot this issue before.\r\nI fi... | null | null |
ranx | AmenRa/ranx | [Feature Request] custom fusion method in optimize_fusion | **Is your feature request related to a problem? Please describe.**
Hi, you’ve done a great job implementing plenty of different fusion algorithms, but I think it will always be a bottleneck.
What would you think about letting the user define their own training function?
**Describe the solution you'd like**
For ex... | https://github.com/AmenRa/ranx/issues/32 | 1,669,630,438,000 | 1,669,641,432,000 | 1 | 1 | A | null | Unimplemented Feature Gap | [
"Hi Paul, \r\n\r\nAre you referring to this passage?\r\n`Finally, there are a few more details of exactly how to combine BM25 and DPR scores worth exploring. As a baseline, we tried using the raw scores directly in the linear combination (exactly as above). However, we noticed that the range of scores from DPR and ... | null | null |
ranx | AmenRa/ranx | [Feature Request] optimize norm and method in optimize_fusion | **Is your feature request related to a problem? Please describe.**
Instead of manually trying every possible fusion techniques, it’d be nice to be able to grid-search all of them, as `optimize_fusion` is already doing for fusion hyperparameters (e.g. weights in `wsum`).
**Describe the solution you'd like**
Allow t... | https://github.com/AmenRa/ranx/issues/33 | 1,669,649,259,000 | 1,669,655,121,000 | null | null | null | null | null | [
"I am aware of that, but I think the users should be aware of the normalization strategies and fusion methods they are using.\r\nTherefore, I do not plan to add this feature any time soon."
] | [
"enhancement"
] | [] |
ranx | AmenRa/ranx | [BUG] Could not find a version that satisfies the requirement ranx | **Describe the bug**
Could not find a version that satisfies the requirement ranx (`pip3 install ranx`)
```
Distributor ID: Ubuntu
Description: Ubuntu 18.04.6 LTS
Release: 18.04
Codename: bionic
pip 21.3.1 from /usr/local/lib/python3.6/dist-packages/pip (python 3.6)
```
Error message:
```
ERROR: Could ... | https://github.com/AmenRa/ranx/issues/34 | 1,669,916,152,000 | 1,669,917,046,000 | 1 | 0 | A | 2 | Documentation Deficiency | [
"This is not a bug, `ranx` requires Python >= 3.7.",
"> This is not a bug, `ranx` requires Python >= 3.7.\r\n\r\nSorry, I could not find this requirement in the documentation.",
"No problem.\r\nPython 3.7 is from mid-2018, that's why I did not write it down.\r\nYou can find all the requirements and dependencies... | null | null |
ranx | AmenRa/ranx | ValueError: max() arg is an empty sequence | Hello,
I'd like to determine what query is causing the following error and how to get around it:
```python
Traceback (most recent call last):
File "main.py", line 43, in perform_tasks
eval(params)
File "main.py", line 25, in eval
eval_helper.perform_eval()
File "/home/celso/projects/XMTC-Baseline... | https://github.com/AmenRa/ranx/issues/35 | 1,673,557,526,000 | 1,674,764,663,000 | 1 | 3 | A | 1 | Validation Gap | [
"I suspect your `filtered_relevance_map` is not valid.\r\nCan you please post a print of it?",
"> I suspect your `filtered_relevance_map` is not valid. Can you please post a print of it?\r\n`filtered_relevance_map`:\r\n.
I thought that one of yo... | https://github.com/AmenRa/ranx/issues/37 | 1,675,422,396,000 | 1,675,431,502,000 | null | null | null | null | null | [
"Hey, I got a pull request for fixing the docs today :D\r\n\r\nStudent's and Fisher's tests very often agree (always?), but Student's is immensely faster to compute.\r\nThat's why I changed the default stat test.\r\n\r\nThe paper you mentioned says the opposite. From the introduction:\r\n\r\n```\r\nStudent’s t, boo... | [] | [] |
ranx | AmenRa/ranx | [Feature Request] Add interpolated recall-precision plot function | **Is your feature request related to a problem? Please describe.**
First of all: This is a really nice library! It helps a lot!
My request is regarding a recall-precision graphic. When I read TREC-related papers, they very often used the interpolated precision-recall-plot to visualize the performance of IR-Systems wh... | https://github.com/AmenRa/ranx/issues/38 | 1,676,289,654,000 | 1,682,961,184,000 | 1 | 4 | A | 1 | Unimplemented Feature Gap | [
"Hi, thanks for the very detailed suggestion/request!\r\nI think this would be a nice addition to the library!\r\nI will take a look to the linked documents in the coming days.",
"I found a really nice lecture on youtube, which explains the recall-precision plot really nice. It can be found here:\r\n\r\nhttps://y... | null | null |
ranx | AmenRa/ranx | [BUG] Missing results causes AssertionError | Missing results causes AssertionError: Qrels and Run query ids do not match
You should make `check_keys` optional in evaluate because sometimes queries do not return any results for lexical-based systems.
https://github.com/AmenRa/ranx/blob/da0aa5246513acb87ec7ac7f516194fea464a2c1/ranx/meta/evaluate.py#L128 | https://github.com/AmenRa/ranx/issues/39 | 1,681,504,378,000 | 1,681,725,582,000 | 1 | 3 | A | 1 | Validation Gap | [
"Hi, I've added the `make_comparable` function to `Run` in v0.3.7.\r\n\r\n```python\r\nrun.make_comparable(qrels)\r\n```\r\n\r\nThe `run` will be modified based on the `qrels` passed as an argument.\r\nQuery results not appearing in `qrels` will be removed from `run` to avoid unpredictable behavior or other bugs.\r... | null | null |
ranx | AmenRa/ranx | feature request: save Report.comparisons as JSON | Hi,
It’d be nice to be able to save a Report comparisons as a JSON file.
However, since it uses frozenset as keys, it is not JSON serializable.
Maybe you could add a method in https://github.com/AmenRa/ranx/blob/master/ranx/frozenset_dict.py to convert the `_map` to a JSON serializable dict, i.e. with `str` keys... | https://github.com/AmenRa/ranx/issues/4 | 1,641,819,478,000 | 1,643,804,329,000 | 1 | 4 | A | 5 | Validation Gap | [
"Hi, an export option for the Report class is already on my to-do list! :)\r\n\r\nI will come back with a proposal so that we can discuss it before I implement the functionality.",
"Hey, sorry for the delay.\r\n\r\nThis is my proposal for the `Report.to_dict` function (I can add a `Report.save_as_json` function f... | null | null |
ranx | AmenRa/ranx | PSP@k: Propensity-scored precision at k | I want to implement the propensity-scored precision at k (PSP@k) as defined above:
$PSP@k = \frac{1}{k} \sum \frac{y_i}{p_i}$
where $p_i$ is the propensity of $y_i$ and $1 \leq i \leq k$.
Therefore, how could I integrate this metric in ranx?
**References**:
[1] Zhang, J., Chang, W.C., Yu, H.F. and Dhillo... | https://github.com/AmenRa/ranx/issues/40 | 1,684,531,436,000 | 1,693,214,941,000 | 1 | 3 | A | 1 | Unimplemented Feature Gap | [
"Hi, sorry for the delay.\r\n\r\nIs this metric similar to reciprocal rank except you average over the top k positions instead of considering only the first retrieved relevant?\r\nCould you please provide an example?",
"See below an example from [pyxclib](https://github.com/kunaldahiya/pyxclib/blob/ae5410f1008074... | null | null |
ranx | AmenRa/ranx | [Feature Request] Expose DCG as metric |
In industry DCG (in both formulations) is a standard and widely used metric. I see it is already implemented as part of NDCG. Is it possible to expose it to the users as a real metric?
| https://github.com/AmenRa/ranx/issues/41 | 1,686,797,462,000 | 1,686,902,292,000 | 1 | 3 | A | 1 | Unimplemented Feature Gap | [
"Hi, `dcg` is now available among the other metrics.\r\nPlease, update your `ranx` installation to `v0.3.13` to use it.\r\n\r\nIf you use `ranx` to evaluate results for your scientific publication, please remember to cite the related papers. Also, `GitHub` stars are much appreciated.\r\n\r\nThank you,\r\n\r\nElias"... | null | null |
ranx | AmenRa/ranx | [BUG] dcg and dcg_burges do not work in the compare function | **Describe the bug**
when adding the newly available `dcg` or `dcg_burges` metric in the compare function I get this error:
```
report = compare(
qrels=qrels,
runs=runs,
metrics=["recall@10","ndcg","rbp.90","rbp.50","dcg_burges"],
max_p=0.05, # P-value threshold
stat_test='student'
)
... | https://github.com/AmenRa/ranx/issues/42 | 1,687,478,029,000 | 1,687,595,854,000 | 1 | 3 | A | 1 | Validation Gap | [
"I think metric labels are missing here: https://github.com/AmenRa/ranx/blob/cf7b00c829a223c988c080ab39562637527346ca/ranx/data_structures/report.py#L24C38-L24C38",
"Yep, I forget about them. Thanks for pointing it out!\r\nFixed in `v0.3.14`"
] | null | null |
ranx | AmenRa/ranx | [Feature Request] Use black to indent the code | **Is your feature request related to a problem? Please describe.**
`black` https://pypi.org/project/black/ is the Uncompromising Code Formatter. You can install it, run `black` and it will format the code properly. How about using it for the project in order to make sure style is consistent? | https://github.com/AmenRa/ranx/issues/44 | 1,687,605,655,000 | 1,687,767,246,000 | null | null | null | null | null | [
"Hi Diego, the code was already `black` formatted.\r\n\r\nThe inconsistency was because I initially forced 80 chars line length in my IDE.\r\nThen, I reverted it to 88 as suggested in `black` documentation but forgot to run `black` project-wise.\r\n\r\nNow code formatting should be consistent."
] | [
"enhancement"
] | [] |
ranx | AmenRa/ranx | [BUG] RBP with multiple relevance levels | **Describe the bug**
Given the formula of RBP of
```
\operatorname{RBP} = (1 - p) \cdot \sum_{i=1}^{d}{r_i \cdot p^{i - 1}}
```
where r_i is the reward/utility, RBP should support multiple relevance levels similar to DCG such that if max relevance level is 2 max rbp value should be 2
| https://github.com/AmenRa/ranx/issues/45 | 1,688,168,471,000 | 1,689,700,603,000 | 1 | 3 | A | 1 | Algorithmic Error | [
"Hi, thanks for pointing it out.\r\n\r\nI suspect I assumed it was a binary relevance metric because the example in the original paper does not account for graded relevance.\r\n\r\nI will fix it in the coming days.",
"Looking at other implementations of RBP I also saw an option to normalize back to [0,1] by divid... | null | null |
ranx | AmenRa/ranx | [Feature Request] Support gzipped files? | **Is your feature request related to a problem? Please describe.**
trec files can be several megabytes, for example `run_*.trec` used for the examples are all more than 20Mb, but once compressed they become less than 10. That would make downloads faster and also loading the file in memory.
**Describe the solution ... | https://github.com/AmenRa/ranx/issues/47 | 1,688,808,334,000 | 1,691,077,797,000 | 1 | 1 | B | 1 | Unimplemented Feature Gap | [
"Both the suggestions sound good to me.\r\n\r\nYou can already save imported runs as highly compressed as `lz4` files with `run.save(\"save/path/run.lz4\")`. Under the hood they are `JSON` files.\r\nI did several tests a few months ago and they should be smaller and faster to extract than gzipped TREC files.\r\n\r\... | null | null |
ranx | AmenRa/ranx | [Feature Request] memory issue / make Run more efficient | Hi Elias,
**Is your feature request related to a problem? Please describe.**
I've noticed that `Run` (and I guess also `Qrels`) consume a lot of memory (RAM) compared to standard python `dict`, e.g. a few GB instead of a few 100s of MB. This gets problematic for somewhat large datasets (e.g. 1M queries)
**Descri... | https://github.com/AmenRa/ranx/issues/48 | 1,688,985,174,000 | 1,689,700,297,000 | 1 | 1 | A | 1 | Resource Mishandling | [
"Dear Paul, \r\n\r\nThe issue is probably due to `numba` (I cannot do anything about that) and a [forced conversion](https://github.com/AmenRa/ranx/blob/7a55106c238d7e147ceb93d0201285b946ef366e/ranx/data_structures/common.py#L53) of the strings used as ids to `numba.types.unicode_type` that I introduced to avoid er... | null | null |
ranx | AmenRa/ranx | question: relevance scores in Qrels | Why is the relevance score mandatory in [Qrels](https://github.com/AmenRa/ranx/blob/master/ranx/qrels.py)?
I don’t see where you use it to compute the metrics https://github.com/AmenRa/ranx/blob/master/ranx/metrics.py
Is there any way to make it optional? If not, would a filler value like `0` for all documents be... | https://github.com/AmenRa/ranx/issues/5 | 1,641,820,146,000 | 1,641,821,005,000 | 1 | 1 | C | null | Documentation Deficiency | [
"Hi, if you only have a list of relevant document ids but no relevance judgment, you should use `1` as relevance judgment score.\r\n\r\n`0` relevance judgment scores are removed before computing metric scores to prevent miscomputations.\r\n\r\nAmong the implemented metrics only NDCG uses the relevance judgment scor... | null | null |
ranx | AmenRa/ranx | Incorrect result for f1 score | Using f1 or _f1_parallel for all qrels and run gives incorrect output. But if I use _f1 on individual query case, it gives correct F1 score.
using below 2 functions return 0 for 4 cases. Ideally it should only be 0 for 1 of all the 18 cases in _qrels & _run passed.
```
from ranx.metrics.f1 import _f1_parallel, _... | https://github.com/AmenRa/ranx/issues/50 | 1,690,473,152,000 | 1,691,077,714,000 | 1 | 3 | A | 1 | Validation Gap | [
"Hi, it is very strange that `_f1` called on every query separately and `_f1_parallel` give different results.\r\n`_f1_parallel` is just a parallelized loop over `_f1` and the queries.\r\n\r\nCan you provide me your `qrels` and `run`?",
"Hi, \r\nthanks for the quick response. \r\nHave attached the files here.\r\n... | null | null |
ranx | AmenRa/ranx | Zero-scored documents | @mponza found a bug when computing recall and promise to document it next week, adding a reminder here for him. Something related to multiple documents having score zero. | https://github.com/AmenRa/ranx/issues/51 | 1,692,982,871,000 | 1,693,647,830,000 | 1 | 3 | A | 1 | Algorithmic Error | [
"**Describe the bug**\r\n\r\nIf a model assigns to a set of documents zero scores, you can still have Recall@k = 1. I am not sure this behaviour is expected or not. \r\n\r\n\r\n**To reproduce**\r\n\r\n```python\r\nfrom ranx import Qrels, Run, evaluate\r\n\r\nqrels_dict = {\"q_1\": {\"d_0\": 1, \"d_1\": 1}}\r\n\r\nr... | null | null |
ranx | AmenRa/ranx | [BUG] Misleading exception message on dataframe types | **Describe the bug**
I'm using the library for the first time with a Pandas dataframe and ran into an exception that was misleading.
**To Reproduce**
Steps to reproduce the behavior:
1. Create a dataframe where the `id` column is of type `int64` e.g. `df['id'] = df.index + 1`
2. Create the qrel like this:
```... | https://github.com/AmenRa/ranx/issues/52 | 1,694,565,334,000 | 1,695,812,431,000 | 1 | 1 | B | 1 | Validation Gap | [
"Hi, sorry for that!\r\nI probably copy-pasted or duplicated lines there.\r\nI will fix it in the next release.",
"Fixed in `v0.3.17`.\r\n\r\nPlease, give `ranx` a star if you haven't yet."
] | null | null |
ranx | AmenRa/ranx | [BUG] Issues when storing/loading Qrels from a dataframe and a parquet file. | **Describe the bug**
Bug when reconstructing Qrels from a pandas dataframe. This bug affects also when reading a Qrel from a parquet file as the pandas to Qrels is used internally.
Pandas version: 1.5.2
Ranx: last pip version
**To Reproduce**
Code:
```python
from ranx import Qrels
qrels = Qrels({'1':{'1':... | https://github.com/AmenRa/ranx/issues/53 | 1,694,599,266,000 | 1,695,983,091,000 | 1 | 1 | C | 1 | Validation Gap | [
"Hi, thanks for reporting this issue.\r\nI will fix it in the next release.",
"Hi, I cannot reproduce.\r\n\r\n`int64` is standard Python `int`.\r\n\r\nWith `ranx`:\r\n```python\r\nfrom ranx import Qrels\r\n\r\nqrels = Qrels({\"1\": {\"1\": 1}, \"2\": {\"2\": 1}})\r\ndf = qrels.to_dataframe()\r\n\r\ndf.dtypes[\"sc... | null | null |
ranx | AmenRa/ranx | [Feature Request] Run.from_df and Run.from_parquet does not allow specifying run name | **Is your feature request related to a problem? Please describe.**
I'm using Pandas dataframes, and comparing different embedding models. I'd like to be able to name my runs so when I compare them, the report shows something other than `run_1`, `run_2`.
**Describe the solution you'd like**
Allow `name` as a named ... | https://github.com/AmenRa/ranx/issues/54 | 1,695,755,230,000 | 1,695,812,443,000 | 1 | 1 | A | null | Unimplemented Feature Gap | [
"Added in `v0.3.17`.\r\n\r\nPlease, give `ranx` a star if you haven't yet."
] | null | null |
ranx | AmenRa/ranx | Question on rank aggregation usage | Thanks for your amazing work. I am very interested in this framework and try to use it to solve my rank aggregation problems. However, I am a little confused about the usage.
For example, I have scores for several items under different ranking rules, as follows:
```
item | rank1 | rank2 | rank3
item1 | 0.4 | 0.8 ... | https://github.com/AmenRa/ranx/issues/55 | 1,696,853,745,000 | 1,696,856,901,000 | null | null | null | null | null | [
"Hi, you can find an example on [Colab](https://colab.research.google.com/github/AmenRa/ranx/blob/master/notebooks/5_fusion.ipynb) and other information [here](https://amenra.github.io/ranx/fusion/).\r\n\r\nAs for your example, you could to something like this:\r\n\r\n```python\r\nfrom ranx import Run, fuse\r\n\r\n... | [] | [] |
ranx | AmenRa/ranx | Getting "Segmentation fault (core dumped)" error | Hello,
Thank you for your amazing work. I am trying to use supervised fusion methods like this:
```python
best_params = optimize_fusion(
qrels=qrels,
runs=[run_4, run_5],
norm='min-max', # Default value
method='posfuse',
metric="mrr@10",
)
combined_run = fuse(
runs=[run_4, run_5],
norm='min-max', # ... | https://github.com/AmenRa/ranx/issues/56 | 1,696,966,605,000 | 1,701,158,002,000 | 1 | 2 | A | 1 | Environment Incompatibility | [
"Hi and thanks for the kind words!\r\n\r\nCan you tell me the OS you are using and the version of `numba` installed?\r\nThe issue is probably related to `numba`, which is a library `ranx` heavily relies on.",
"\r\nAt first I was using an Ubuntu remote server (Ubuntu 20.04.6 LTS). I was getting that error for mos... | null | null |
ranx | AmenRa/ranx | [Feature Request] stddev statistic | It would be a good to additionally calculate the sample variance within the average of the metrics.
Like `ndcg@50: 11, stddev: 2.8`
| https://github.com/AmenRa/ranx/issues/57 | 1,697,116,316,000 | 1,701,159,671,000 | 1 | 3 | B | 2 | Unimplemented Feature Gap | [
"Hi! Yes, I should add it.\r\n\r\nIn the mean time, you can compute it as follows:\r\n```python\r\nfrom ranx import Qrels, Run, evaluate\r\nimport numpy as np\r\n\r\nqrels_dict = { \"q_1\": { \"d_12\": 5, \"d_25\": 3 },\r\n \"q_2\": { \"d_11\": 6, \"d_22\": 1 } }\r\n\r\nrun_dict = { \"q_1\": { \"d_12\... | null | null |
ranx | AmenRa/ranx | Couldn't find any documentation about Qrel and run score range | Hi,
In the current api documentation I see some examples of how to structure `qrel` and `run` dicts, but there is no detail about what should be the range of `qrel` relevance score, I assume its 0 for non-relevant cases and 1 or more for graded relevance, but is there a upper limit to integer value?
Similarly, f... | https://github.com/AmenRa/ranx/issues/58 | 1,699,176,600,000 | 1,699,517,744,000 | 1 | 1 | B | 1 | Documentation Deficiency | [
"Hi and thanks for your question.\r\n\r\n`qrels` scores can be any integer, positive or negative.\r\nAll the implemented metrics but [`bpref`](https://amenra.github.io/ranx/metrics/#bpref) only considers relevant documents, i.e. discard all `qrels` < 1. You can change this threshold by adding `-ln` (`n` is a parame... | null | null |
ranx | AmenRa/ranx | [Feature Request] Propensity-scored Metrics |
**Is your feature request related to a problem? Please describe.**
In some extremely challenging scenarios for information retrieval, items relevant to a query follow a long tail distribution. Thus, there are a few extremely frequent relevant items (head items) and many extremely rare relevant items (tail items). Si... | https://github.com/AmenRa/ranx/issues/59 | 1,701,705,077,000 | null | 1 | 3 | A | 1 | Unimplemented Feature Gap | [
"Adding these new features to the context of my research would be very beneficial. Are there any updates on their development?",
"Any progress here?",
"Hi, I have no time to work on this at the moment.\r\nYou can implement it by yourself and make a PR if you want.",
"> Hi, I have no time to work on this at th... | null | null |
ranx | AmenRa/ranx | question: Comparing models with multiple runs | First of all, great work on this code. I have been looking for a definitive package to evaluate ranking models and I believe this is that package.
My question is perhaps a bit out of the domain, but it could help others in the future. How would you deal with comparing 2 models where each has multiple runs (e.g., run... | https://github.com/AmenRa/ranx/issues/6 | 1,642,071,407,000 | 1,642,170,773,000 | 1 | 3 | B | 2 | Unimplemented Feature Gap | [
"Hi and thanks for the kind words! :)\r\n\r\nThat's a tricky question.\r\nI actually don't have a _scientific_ answer, but your idea sounds reasonable to me.\r\n\r\nRight now `ranx` does not support what you say out of the box but it can probably be achieved using the \"low level\" function you find in the source c... | null | null |
ranx | AmenRa/ranx | How do we compare different runs with multiple folds per run? | How do we compare different runs with multiple folds per run?
For instance, assume we have 10-folds for `run_1`, ... `run_5`?
```python
from ranx import compare
# Compare different runs and perform Two-sided Paired Student's t-Test
report = compare(
qrels=qrels,
runs=[run_1, run_2, run_3, run_4, run_... | https://github.com/AmenRa/ranx/issues/60 | 1,705,518,315,000 | null | null | null | null | null | null | [
"Also, could you provide a simple explanation of how to interpret the report?\r\n```md\r\n# Model MAP@100 MRR@100 NDCG@10\r\n--- ------- -------- -------- ---------\r\na model_1 0.320ᵇ 0.320ᵇ 0.368ᵇᶜ\r\nb model_2 0.233 0.234 0.239\r\nc model_3 0.308ᵇ 0.309ᵇ 0.... | [] | [] |
ranx | AmenRa/ranx | [Question] About the correction among multiple hypotheses | I came across a point in the [Fuhr](https://www.sigir.org/wp-content/uploads/2018/01/p032.pdf) article where it talks about the need for additional correction for multiple hypotheses test (e.g., Bonferroni correction). Looking at the `paired_student_t_test` function code, it seems like the correction procedure is not t... | https://github.com/AmenRa/ranx/issues/61 | 1,706,173,425,000 | 1,719,856,597,000 | 1 | 3 | B | 2 | Documentation Deficiency | [
"Hi and sorry for the delay.\r\n\r\nYou can apply Bonferroni correction by changing the `max_p` parameter of the `compare` function (Maximum p-value to consider an increment as statistically significant. Defaults to 0.01.).\r\n\r\nOtherwise, you can use the [Tukey's HSD Test](https://www.itl.nist.gov/div898/handboo... | null | null |
ranx | AmenRa/ranx | [Question] How to compute precision for a retriever operating at passage-level | If the retriever is operating at passage / chunk-level and hence the retrieved results can have duplicate IDs as shown below:
Top-10 retrieval list for each of the queries:
```
q_1: ['d_1', 'd_1', 'd_1', 'd_1', 'd_1', 'd_1', 'd_5', 'd_5', 'd_5', 'd_5']
q_2: ['d_4', 'd_4', 'd_4', 'd_4', 'd_2', 'd_2', 'd_6', 'd_6', '... | https://github.com/AmenRa/ranx/issues/63 | 1,709,487,806,000 | 1,719,856,185,000 | 1 | 3 | A | 1 | Documentation Deficiency | [
"Hi and sorry for the delay.\r\n\r\nI suggest you use ids for passages OR perform a pooling operation on the scores for the passages for a given document, e.g. taking the maximum scores.\r\n\r\nHope it helps."
] | null | null |
ranx | AmenRa/ranx | JIT compilation on serverless (i.e. Modal Labs) | I'm noticing that it takes a really long time to start up anything with ranx in a serverless setting, e.g. Modal Labs. I tried adding a step during the build process where I call the functions that need to be cached, but it doesn't seem to help. This is true even if I explicitly set the NUMBA_CACHE_DIR to a directory t... | https://github.com/AmenRa/ranx/issues/64 | 1,710,371,901,000 | null | 1 | 0 | A | 3 | Documentation Deficiency | [
"Hi and sorry for the delay.\r\nUnfortunately, I cannot help with this.\r\nIf you find a solution, please post it here for others.\r\nThanks.",
"@andersonbcdefg @AmenRa please check https://github.com/AmenRa/ranx/issues/75 "
] | null | null |
ranx | AmenRa/ranx | I'm getting NaN for the BPref measurement | Not really sure what more information to provide.
I'm new to ranx (and trec_eval, in general).
I'm loading qrels and results trec files into a Qrels and Run, then simply oing:
measurements = str(evaluate(qrels, run, ["mrr", "bpref", "precision@6", "precision@10", "precision@25", "ndcg", "ndcg@6", "recall@6", "r... | https://github.com/AmenRa/ranx/issues/66 | 1,721,850,519,000 | 1,721,888,130,000 | 1 | 3 | A | 1 | Validation Gap | [
"Ok, this should help: The simple dict example produces it too:\r\n```\r\nfrom ranx import Qrels, Run, evaluate\r\n\r\nqrels_dict = { \"q_1\": { \"d_12\": 5, \"d_25\": 3 },\r\n \"q_2\": { \"d_11\": 6, \"d_22\": 1 } }\r\n\r\nrun_dict = { \"q_1\": { \"d_12\": 0.9, \"d_23\": 0.8, \"d_25\": 0.7,\r\n ... | null | null |
ranx | AmenRa/ranx | [BUG] Run with one empty results error | **Describe the bug**
A Run with only one empty results lead to ValueError
May related to the issue here: https://github.com/AmenRa/ranx/issues/51
**To Reproduce**
run = Run({"q_1": {}})
qrel = Qrels({"q_1": { 'reference_document': 1.0}})
| https://github.com/AmenRa/ranx/issues/67 | 1,725,453,984,000 | null | 1 | 2 | A | 1 | Validation Gap | [] | null | null |
ranx | AmenRa/ranx | feature request: hits (or accuracy?) | Hi,
@osf9018 mentioned it in #2 but I guess it’s better to create a specific issue.
# Motivation
It is often difficult to estimate the total number of relevant document for a query.
For example, in Question Answering, if you have a large enough Knowledge Base, you can find the answer to your question in a surpr... | https://github.com/AmenRa/ranx/issues/7 | 1,642,089,121,000 | 1,643,734,956,000 | 1 | 3 | A | 1 | Configuration Error | [
"Hi, I can add it to the pool of the provided metrics for sure! :)\r\n\r\nI'm just not confident about how I should call it.\r\nCould `success_rate` or `hit_rate` be appropriate?\r\nI can even call it `hits` and rename or hide the current `hits` metric.\r\n\r\nWhat do you think?",
"hit_rate seems fine, as you wan... | null | null |
ranx | AmenRa/ranx | Fisher's Randomization Test yields nondeterministic results | **Describe the bug**
When running Fisher's Randomization Test multiple times using the same data and same parameters, the significance results vary between runs. This happens even when the random seed is fixed and also when just a single thread is used.
**To Reproduce**
Run `ranx.compare(qrels, runs, metrics=["pre... | https://github.com/AmenRa/ranx/issues/70 | 1,732,201,541,000 | null | 1 | 3 | A | 1 | Algorithmic Error | [] | null | null |
ranx | AmenRa/ranx | [Feature Request] Support for non-integer relevance labels | I would like to be able to use real-valued labels, and compute DCG and NDCG metrics that respect the differences in labels that at less than 1. So for example:
- A document with relevance label = 1 would have Burges gain of `2**1 - 1 = 1.0`
- But a document with relevance label = 1.9 would have Burges gain of `2**1.9 -... | https://github.com/AmenRa/ranx/issues/71 | 1,737,721,058,000 | null | 1 | 3 | A | 1 | Unimplemented Feature Gap | [
"+1 for this. Continuous labels are produced from clickstream data, and may be used in some of these metrics."
] | null | null |
ranx | AmenRa/ranx | [BUG] Crash when trying to make a Run comparable if numba JIT is disabled | **Describe the bug**
Ranx is very very slow when I use it with many small Run and Qrels objects (e.g. to do question-by-question comparisons). This is likely because Ranx uses the Numba JIT compiler, which is optimized for few large objects with long compute times. (See previous issue: https://github.com/AmenRa/ranx/i... | https://github.com/AmenRa/ranx/issues/74 | 1,752,057,943,000 | null | 1 | 1 | A | 1 | Validation Gap | [
"Thanks @milyenpabo I’ve also been thinking about supporting Ranx without Numba, by implementing the metrics using NumPy vectorized operations. I’d be curious to implement a parallel version of the metrics and then benchmark both Numba and non-Numba approaches across different collection sizes. My gut feeling is th... | null | null |
ranx | AmenRa/ranx | [Feature Request] Make Numba Optional with Pure NumPy Fallbacks | Just pitching this idea to see if we would like to try this, also happy to develop it in a dev branch to see how we feel about it. But first I'd like to get your feelings about this strategy @AmenRa @milyenpabo and @andersonbcdefg.
## Problem
Numba is currently a hard dependency that significantly impacts the user ex... | https://github.com/AmenRa/ranx/issues/75 | 1,754,514,651,000 | null | null | null | null | null | null | [
"This seems like a good idea to me. After opening the issue you mentioned, I ended up just re-implementing the functionality I needed from this library in pure python or numpy.",
"I see this may be beneficial for some use cases. However, it would be wise to first evaluate when this is the case before implementing... | [
"enhancement"
] | [
"https://github.com/AmenRa/ranx/issues/64",
"https://github.com/AmenRa/ranx/pull/76"
] |
End of preview. Expand in Data Studio
EvalEng Dataset
GitHub issues collected from 57 ML evaluation harnesses, annotated with workflow stage, root cause, and evaluation strategy labels.
Dataset Details
- Paper: Towards Evaluation Engineering
- Repository: https://github.com/zhimin-z/EvalEng
- License: MIT
Fields
| Field | Type | Description |
|---|---|---|
harness_name |
string | Name of the ML evaluation harness |
github_repo |
string | GitHub repository (owner/repo) |
issue_title |
string | Issue title |
issue_body |
string | Issue body text |
issue_url |
string | Direct URL to the GitHub issue |
issue_created_at |
int64 | Creation timestamp (ms since epoch) |
issue_closed_at |
int64 | Closed timestamp (ms since epoch) |
is_related |
float64 | Whether the issue is related to evaluation (1.0 = yes) |
stage |
float64 | Workflow stage label (numeric) |
step |
string | Workflow step label |
strategy |
float64 | Evaluation strategy label (numeric) |
root_cause_label |
string | Root cause category |
issue_comments |
list[string] | Issue comment texts |
issue_labels |
list[string] | GitHub labels on the issue |
issue_cross_referenced |
list[string] | Cross-referenced issue/PR URLs |
Citation
@article{zhao2025evaleng,
title={Towards Evaluation Engineering: An Empirical Study of ML Evaluation Harnesses in the Wild},
author={Zhimin Zhao and Zehao Wang and Abdul Ali Bangash and Bram Adams and Ahmed E. Hassan},
year={2026},
eprint={2605.24213},
archivePrefix={arXiv},
primaryClass={cs.SE},
url={https://arxiv.org/abs/2605.24213},
}
- Downloads last month
- -