snooker-testbed-phase4-attempt1-killed-v1

Phase-4 main training attempt 1, KILLED early per pre-defined fail criterion (score < 1.5 at step 4M with std drift up). Job 7227026, ran 2026-04-26 13:26 → 18:30 UTC, killed at step 4.2M (64m wall). Config: 8M total, Path-2+Path-3 combined (constant +2.0 hit-any-ball bonus + foul_multiplier 1.0→2.5 ramp over steps 1M-4M, n_envs=8, eval_episodes=32). VERDICT: FAIL. Std exploded UPWARD during ramp (0.91 at 800k → 1.55 at 4.2M, monotonic). Score regressed from peak 1.06 (step 2.4M, pre-ramp) to 0.22 at step 4M (full multiplier=2.5). The 2.5× ceiling was too aggressive: drives advantages so negative that PPO's entropy bonus dominates and de-concentrates the policy. Phase 4b will lower the ceiling to 1.5.

Dataset Info

• Rows: 10
• Columns: 11

Columns

Column	Type	Description
step	Value('int64')	Global PPO timestep (every 400k)
curriculum_stage	Value('int64')	Stage at this step (all stage 2)
mean_score	Value('float64')	Snooker points scored, mean over 32 eps. Peak 1.06 pre-ramp, regressed to 0.22 by step 4M.
max_score	Value('float64')	Best episode score (peak 17 at step 2.8M — highest of any run, but isolated)
mean_shots	Value('float64')	Mean shots per episode (200 = truncated)
mean_efficiency	Value('float64')	mean_score / mean_shots
mean_foul_rate	Value('float64')	Stays 98-99% throughout — even with full 2.5× foul multiplier
episodes	Value('int64')	32 eval episodes per eval
source_file	Value('string')	Original metrics_step_*.json
std_at_nearest_ckpt	Value('float64')	policy.log_std exp(value) at nearest checkpoint — KEY DIAGNOSTIC. 0.91 at step 800k (lowest), then climbs to 1.55 at 4.2M.
std_step	Value('int64')	Step number for the std value

Generation Parameters

{
  "script_name": "sbatch/train_sim.sbatch (Phase 4 attempt 1, KILLED at step 4.2M)",
  "model": "stable-baselines3 PPO, MlpPolicy net_arch=[256,256]",
  "description": "Phase-4 main training attempt 1, KILLED early per pre-defined fail criterion (score < 1.5 at step 4M with std drift up). Job 7227026, ran 2026-04-26 13:26 \u2192 18:30 UTC, killed at step 4.2M (64m wall). Config: 8M total, Path-2+Path-3 combined (constant +2.0 hit-any-ball bonus + foul_multiplier 1.0\u21922.5 ramp over steps 1M-4M, n_envs=8, eval_episodes=32). VERDICT: FAIL. Std exploded UPWARD during ramp (0.91 at 800k \u2192 1.55 at 4.2M, monotonic). Score regressed from peak 1.06 (step 2.4M, pre-ramp) to 0.22 at step 4M (full multiplier=2.5). The 2.5\u00d7 ceiling was too aggressive: drives advantages so negative that PPO's entropy bonus dominates and de-concentrates the policy. Phase 4b will lower the ceiling to 1.5.",
  "hyperparameters": {
    "algorithm": "PPO",
    "total_timesteps": 8000000,
    "actual_steps_completed": 4200000,
    "killed": true,
    "kill_reason": "score 0.22 < 1.5 at step 4M with std 1.44 drifting up \u2014 pre-defined fail criterion",
    "n_envs": 8,
    "n_steps": 512,
    "batch_size": 2048,
    "ent_coef": 0.01,
    "gamma": 0.99,
    "curriculum_stages": [
      0,
      1,
      2,
      3,
      4
    ],
    "advancement_threshold": -400.0,
    "curriculum_min_episodes_per_stage": 100,
    "curriculum_window_size": 50,
    "reward_shot_cost": 0.01,
    "reward_pot_bonus": 0.5,
    "reward_completion_bonus": 10.0,
    "reward_position_shaping": 0.0,
    "reward_hit_any_ball_bonus": 2.0,
    "reward_foul_multiplier": 2.5,
    "reward_foul_multiplier_ramp_start_step": 1000000,
    "reward_foul_multiplier_ramp_end_step": 4000000,
    "code_version": "2026-04-24-phase3c",
    "commit": "c7747d6"
  },
  "input_datasets": [],
  "experiment_name": "snooker-testbed",
  "job_id": "torch:7227026",
  "cluster": "torch",
  "artifact_status": "killed",
  "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-phase4-attempt1-killed-v1", split="train")
print(f"Loaded {len(dataset)} rows")

---