openenv.yaml

# openenv.yaml
# ─────────────────────────────────────────────────────────────────────────────
# Manifest for the Data Cleaning Pipeline OpenEnv environment.
#
# Field reference
# ───────────────
# Required by the CLI (serve / build / push / validate):
#   spec_version  — always 1 for this generation of the spec
#   name          — environment identifier used by the CLI and auto-discovery
#   type          — "space" means it can be deployed as a Hugging Face Space
#   runtime       — "fastapi" tells the server how to boot
#   app           — Python import path to the FastAPI app object
#   port          — port the server listens on inside the container
#
# Read by AutoEnv auto-discovery (openenv.auto._discovery):
#   name          — maps to env_key after stripping the "_env" suffix
#   description   — human-readable label shown in env listings
#   spec_version  — stored in EnvironmentInfo for introspection
#   action        — EXPLICIT override of the auto-inferred class name
#   observation   — EXPLICIT override of the auto-inferred class name
#
# NOTE on action / observation overrides:
#   Auto-discovery infers class names from the env name using PascalCase:
#     "data_cleaning_env" → base "data_cleaning" → "CleanAction"
#   Our actual class is named "CleanAction" (not "CleanAction"),
#   so these fields MUST be set to avoid ImportError on AutoEnv.from_env().
#
# All other fields (tasks, reward, tags) are informational.  They are not
# parsed by the current OpenEnv tooling but are preserved in
# EnvironmentInfo.manifest and available to the web UI and external tools.
# ─────────────────────────────────────────────────────────────────────────────

# ── Core deployment fields ────────────────────────────────────────────────────

spec_version: 1
name: data_cleaning_env
type: space
runtime: fastapi
app: server.app:app
port: 8000

# ── Package metadata ──────────────────────────────────────────────────────────

version: "1.0.0"

description: >-
  Data cleaning pipeline: the agent receives a dirty CSV and must detect
  and fix type errors, missing values, outliers, and schema inconsistencies
  to match a hidden ground-truth dataset. Three tasks (easy → medium → hard)
  with a deterministic grader that returns a continuous score in [0.0, 1.0].

# ── Auto-discovery class overrides ───────────────────────────────────────────
# These override auto-inferred names (which would be CleanAction /
# CleanAction) to match the actual class names defined in models.py.

action: CleanAction
observation: CleanObservation

# The client class is correctly inferred as DataCleaningEnv (data_cleaning →
# DataCleaning + Env), which matches client.py, so no override is needed.

# ── Tags (informational) ──────────────────────────────────────────────────────

tags:
  - data-cleaning
  - tabular
  - real-world
  - hackathon

# ── Task manifest (informational) ─────────────────────────────────────────────
# One entry per task. These values mirror the constants in models.py
# (MAX_STEPS, DONE_THRESHOLD) and the descriptions in dataset_factory.py.

tasks:
  - id: easy
    name: Fix obvious errors
    description: >-
      50-row sales CSV with 29 injected dirty cells: 10 type mismatches
      (text in numeric columns), 8 missing values, 5 far-future dates
      (year 2099), and 6 cells with leading/trailing whitespace.
      Graded by exact cell-level match against the ground truth (0.0–1.0).
    dataset_rows: 50
    dirty_cells: 29
    max_steps: 40
    done_threshold: 0.95

  - id: medium
    name: Outlier detection without false positives
    description: >-
      200-row customer transaction CSV with 15 true statistical outliers
      (negative or > $2000 amounts) that must be fixed or removed, 5 valid
      large transactions ($900–$2000) that must NOT be removed, and 12
      category spelling typos. Graded by F1 score on outlier detection
      (0.5 weight) and typo correction rate (0.5 weight).
    dataset_rows: 200
    dirty_cells: 27
    max_steps: 80
    done_threshold: 0.85

  - id: hard
    name: Multi-source schema normalisation and deduplication
    description: >-
      430-row CSV (400 clean + 30 duplicates) merged from 3 fictional data
      sources with inconsistent column naming (e.g. cust_id / customer_id /
      CustomerID), mixed date formats (ISO, US, EU), and ~30 duplicate rows
      (exact and near-duplicate). Agent must infer the canonical 9-column
      schema without explicit documentation. Graded by schema match (40%),
      deduplication F1 (35%), and date format compliance (25%).
    dataset_rows: 430
    canonical_rows: 400
    canonical_columns: 9
    duplicate_rows: 30
    max_steps: 150
    done_threshold: 0.80

# ── Reward function summary (informational) ───────────────────────────────────

reward:
  type: dense
  range: [-0.5, 1.0]
  step_cost: -0.005
  components:
    - name: progress
      weight: primary
      description: >-
        Grader score delta each step (curr_score − prev_score).
        The main learning signal — any cell fixed produces a non-zero reward.

    - name: efficiency_bonus
      weight: "+0.10 × (1 − step_fraction)"
      description: >-
        Small bonus awarded the step the episode is solved (score crosses
        done_threshold). Rewards finishing early relative to the step budget.

    - name: false_positive_penalty
      weight: -0.15
      description: >-
        Applied when DROP_ROW removes a valid-extreme row in the medium task.
        Penalises aggressive deletion without checking schema_hint.

    - name: early_done_penalty
      weight: -0.20
      description: >-
        Applied when the agent sends DONE with current_score < 0.60.
        Discourages giving up prematurely.

    - name: step_cost
      weight: -0.005
      description: >-
        Fixed cost every step regardless of outcome.
        Prevents infinite loops and padding.