snooker-testbed-legacy-ppo-v1

Legacy snooker PPO training run — pre-refactor. 9 slurm jobs between 2026-04-15 and 2026-04-17 on torch (h200). Last job 6428636 was SIGTERM'd at ~4.02M / 10M PPO timesteps. Policy did NOT learn to play: mean score 0-8 (out of 147 max), foul rate 95-100% throughout training. Preserved for ablation comparison against the forthcoming Phase-2 refactor.

Dataset Info

• Rows: 321
• Columns: 8

Columns

Column	Type	Description
step	Value('int64')	Global PPO timestep at which the eval was run
mean_score	Value('float64')	Mean snooker points scored across the eval episodes (max possible 147 per episode)
max_score	Value('float64')	Max snooker points scored in any single eval episode
mean_shots	Value('float64')	Mean number of shots taken per episode (capped at max_shots=200)
mean_efficiency	Value('float64')	mean_score / mean_shots — very low values indicate most shots were unproductive
mean_foul_rate	Value('float64')	Fraction of shots that were fouls in this eval (0-1)
episodes	Value('int64')	Number of eval episodes aggregated (legacy runs used only 2, which is the root cause of high variance)
source_file	Value('string')	Name of the original metrics_step_*.json file on torch

Generation Parameters

{
  "script_name": "snooker.main --mode train --algorithm PPO",
  "model": "stable-baselines3 PPO, MlpPolicy net_arch=[256,256]",
  "description": "Legacy snooker PPO training run \u2014 pre-refactor. 9 slurm jobs between 2026-04-15 and 2026-04-17 on torch (h200). Last job 6428636 was SIGTERM'd at ~4.02M / 10M PPO timesteps. Policy did NOT learn to play: mean score 0-8 (out of 147 max), foul rate 95-100% throughout training. Preserved for ablation comparison against the forthcoming Phase-2 refactor.",
  "hyperparameters": {
    "algorithm": "PPO",
    "n_envs": 8,
    "n_steps": 512,
    "batch_size": 2048,
    "learning_rate": 0.0003,
    "ent_coef": 0.01,
    "gamma": 0.99,
    "curriculum_stages": [
      1,
      2,
      3,
      4
    ],
    "advancement_threshold": 5.0,
    "eval_episodes": 2,
    "eval_interval_steps": 25000,
    "reward_shot_cost": 0.1,
    "reward_position_shaping": 0.1,
    "reward_pot_bonus": 0.5,
    "reward_completion_bonus": 10.0,
    "action_dim": 4,
    "obs_dim": 73
  },
  "input_datasets": [],
  "experiment_name": "snooker-testbed",
  "job_id": "empire:6428636",
  "cluster": "torch",
  "artifact_status": "final",
  "canary": false
}

Usage

from datasets import load_dataset

dataset = load_dataset("aditijc/snooker-testbed-legacy-ppo-v1", split="train")
print(f"Loaded {len(dataset)} rows")

---