qwenjev / eval_report_random_baseline_partial.json
tchbcb's picture
Upload eval_report_random_baseline_partial.json with huggingface_hub
e25a546 verified
Raw History Blame Contribute Delete
2.48 kB
{
"meta": {
"backend": "random-uniform",
"seed_note": "rng = Random(42 + idx*100) per run, fixed for reproducibility",
"games": [
"ar25@600",
"ar25@150",
"bp35@150"
],
"max_actions": [
600,
150,
150
],
"timestamp": "2026-09-24 05:01:27",
"note": "Random-baseline control. T4 8K9ITI8A was timeout-recycled (502) MID-RUN: ar25@600 and ar25@150 completed and printed RESULT lines (captured verbatim in the session log); bp35@150 was lost and the report file was never written (written only at end of runfile). Entries below transcribe the two captured RESULT lines; bp35@150 marked null = pending re-run on next T4 (1-minute job)."
},
"games": [
{
"game_id": "ar25",
"budget": 600,
"win": false,
"final_state": "NOT_FINISHED",
"best_level": 0,
"resets": 6,
"actions": 594,
"seconds": 0.8,
"action_hist": {
"ACTION6": 88,
"ACTION1": 95,
"ACTION3": 77,
"ACTION2": 97,
"ACTION5": 83,
"ACTION4": 71,
"ACTION7": 83
}
},
{
"game_id": "ar25",
"budget": 150,
"win": false,
"final_state": "NOT_FINISHED",
"best_level": 0,
"resets": 1,
"actions": 149,
"seconds": 0.2,
"action_hist": {
"ACTION5": 19,
"ACTION6": 20,
"ACTION4": 14,
"ACTION2": 23,
"ACTION3": 21,
"ACTION7": 28,
"ACTION1": 24
}
},
{
"game_id": "bp35",
"budget": 150,
"status": "lost_to_t4_recycle",
"best_level": null,
"resets": null,
"actions": null
}
],
"summary": {
"played": 2,
"wins": 0,
"best_level_sum": 0,
"resets_sum": 7,
"actions_sum": 743,
"pending": [
"bp35@150"
]
},
"bp35_150": {
"game_id": "bp35",
"win": false,
"final_state": "NOT_FINISHED",
"best_level": 0,
"resets": 2,
"actions": 148,
"seconds": 0.3,
"action_hist": {
"ACTION3": 39,
"ACTION6": 37,
"ACTION4": 41,
"ACTION7": 31
},
"note": "completed on 0S3BN0UN 2026-09-24 (was lost_to_t4_recycle)"
},
"status": "complete",
"conclusion": "Random baseline 0 levels across all 3 configs (ar25@600: 0W/6R, ar25@150: 0W/1R, bp35@150: 0W/2R) -> 0-level = noise floor for <=600-step budgets. Model 0-reset behavior = learned safety, not failure. Confirms v5 agent-teaching direction."
}