Hugging Face's logo

Datasets:
aditijc
/
snooker-testbed-phase5-main-v1

Modalities:
Tabular
Text
Formats:
parquet
Size:
< 1K
Tags:
snooker-testbed
phase-5
main-run
ppo
multidiscrete
foul-ramp
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
step
int64
250k
5M
curriculum_stage
int64
2
2
mean_score
float64
0.63
8.75
max_score
float64
8
28
mean_shots
float64
200
200
mean_efficiency
float64
0
0.04
mean_foul_rate
float64
0.89
0.98
episodes
int64
16
16
250,000
2
0.625
8
200
0.003125
0.975313
16
500,000
2
3.875
13
200
0.019375
0.967188
16
750,000
2
5.6875
17
200
0.028438
0.94875
16
1,000,000
2
2.625
13
200
0.013125
0.960312
16
1,250,000
2
6.3125
28
200
0.031563
0.945938
16
1,500,000
2
7.3125
19
200
0.036563
0.94625
16
1,750,000
2
6.3125
24
200
0.031563
0.917812
16
2,000,000
2
3.1875
16
200
0.015938
0.95125
16
2,250,000
2
3.375
21
200
0.016875
0.937188
16
2,500,000
2
8
20
200
0.04
0.916563
16
2,750,000
2
6.5625
23
200
0.032813
0.939062
16
3,000,000
2
6.4375
17
200
0.032188
0.916875
16
3,250,000
2
5.8125
20
200
0.029063
0.915938
16
3,500,000
2
8.75
22
200
0.04375
0.93875
16
3,750,000
2
4.125
17
200
0.020625
0.948125
16
4,000,000
2
4.8125
22
200
0.024063
0.929688
16
4,250,000
2
5.625
17
200
0.028125
0.895625
16
4,500,000
2
5.1875
14
200
0.025938
0.893125
16
4,750,000
2
6.125
19
200
0.030625
0.931563
16
5,000,000
2
7
19
200
0.035
0.922187
16

snooker-testbed-phase5-main-v1

PHASE 5 MAIN RUN. Combines the 4F2 breakthrough recipe (PPO MultiDiscrete + augmented obs + curriculum + stage-0 bootstrap) with foul-multiplier ramp 1.0→1.5 over steps 1M-3M. Job 7320872 on torch h200, 1h20m wall (5M steps in single slot, no requeue needed), ran 2026-04-27 20:20-21:40 UTC. RESULTS: mean across 20 evals: 5.39 (target was > 5). Peak score 8.75 at step 3.5M (target was > 10 — close, single-ep max 28 at 1.25M is highest ever). Foul rate floor 89.3% at step 4.5M (target was < 85% — close). The foul-multiplier ramp clearly worked: pre-ramp foul ~95-97%, post-ramp dropped to 89-93% range. Score continued climbing in late phase. Policy genuinely plays snooker now.

Dataset Info

  • Rows: 20
  • Columns: 8

Columns

Column Type Description
step Value('int64') Global PPO timestep (250k-5M)
curriculum_stage Value('int64') No description provided
mean_score Value('float64') Mean over 16 eps. Sustained climb. Peak 8.75 at step 3.5M.
max_score Value('float64') Best single-ep score. Peak 28 at step 1.25M (highest of any run).
mean_shots Value('float64') No description provided
mean_efficiency Value('float64') No description provided
mean_foul_rate Value('float64') Foul rate floor 89.3% at step 4.5M — first run < 90%.
episodes Value('int64') No description provided

Generation Parameters 

{
  "script_name": "sbatch/train_sim.sbatch (Phase 5 \u2014 5M steps + foul ramp 1.0\u21921.5)",
  "model": "stable-baselines3 PPO, MlpPolicy net_arch=[256,256], MultiDiscrete([36,8,5,5])",
  "description": "PHASE 5 MAIN RUN. Combines the 4F2 breakthrough recipe (PPO MultiDiscrete + augmented obs + curriculum + stage-0 bootstrap) with foul-multiplier ramp 1.0\u21921.5 over steps 1M-3M. Job 7320872 on torch h200, 1h20m wall (5M steps in single slot, no requeue needed), ran 2026-04-27 20:20-21:40 UTC. RESULTS: mean across 20 evals: 5.39 (target was > 5). Peak score 8.75 at step 3.5M (target was > 10 \u2014 close, single-ep max 28 at 1.25M is highest ever). Foul rate floor 89.3% at step 4.5M (target was < 85% \u2014 close). The foul-multiplier ramp clearly worked: pre-ramp foul ~95-97%, post-ramp dropped to 89-93% range. Score continued climbing in late phase. Policy genuinely plays snooker now.",
  "hyperparameters": {
    "algorithm": "PPO",
    "total_timesteps": 5000000,
    "n_envs": 8,
    "n_steps": 512,
    "batch_size": 2048,
    "ent_coef": 0.01,
    "learning_rate": 0.0003,
    "gamma": 0.99,
    "discrete_actions": true,
    "discrete_phi_bins": 36,
    "discrete_force_bins": 8,
    "discrete_spin_bins": 5,
    "obs_dim": 79,
    "reward_foul_multiplier": 1.5,
    "reward_foul_multiplier_ramp_start_step": 1000000,
    "reward_foul_multiplier_ramp_end_step": 3000000,
    "code_version": "2026-04-27-phase4f-discrete",
    "commit": "877d785"
  },
  "input_datasets": [],
  "experiment_name": "snooker-testbed",
  "job_id": "torch:7320872",
  "cluster": "torch",
  "artifact_status": "final",
  "canary": false
}

Experiment Documentation

For complete experiment details, see https://github.com/aditijc/snooker-testbed

Usage 

from datasets import load_dataset

dataset = load_dataset("aditijc/snooker-testbed-phase5-main-v1", split="train")
print(f"Loaded {len(dataset)} rows")
Downloads last month
66
Number of rows:
20
Total file size:
7.52 MB