rh-dynamics-runs

Raw logs and data of the training runs behind the reward hacking dynamics LoRA adapters. Checkpoints are not included (the adapters are in the model repos).

Layout: <model repo name>/seed-<rollout seed>/

Path	Content
rollouts/<step>.jsonl	every training rollout of the step (16 prompts x 8): input, output, score, gts, step, uid
validation/<step>.jsonl	held-out set (119 problems, 1 sample each) with full grading fields (eq_correct, defined_test_func, reward_hack_strict, reward_hack_loose, task_truncated, ...)
rollout_labels.jsonl	training rollouts re-graded with scripts/label_rollouts.py (one line per rollout, in file order)
train.log, train_full_stdout.log	driver stdout with per-step metrics; the second also has the launcher output
agent_loops.yaml	agent-loop config of the run
wandb/	offline wandb run directory
launch/	launch script / command used

Runs

Path	Hardware	Notes
Qwen3.5-4B-rh-dynamics/seed-42	4 × H200 (Slurm)	hacking took off at step ~42
Qwen3.5-4B-rh-dynamics/seed-1	2 × B200	no hacking
Qwen3.5-4B-rh-dynamics/seed-2	2 × B200	no hacking
Qwen3.5-4B-rh-dynamics/seed-3	2 × B200	no hacking; computation-preserving speed settings (no offload, no gradient checkpointing, chunked log-prob)
Qwen3.5-4B-rh-dynamics/seed-4	2 × B200	no hacking; same settings as seed 3
Qwen3.5-2B-rh-dynamics/seed-42	2 × H200	no hacking; thinking mode switched on (enable_thinking=True, the 2B template defaults to non-thinking); same speed settings as 4B seed 3
Qwen3.5-2B-rh-dynamics/seed-1	2 × H200	no hacking; same settings as 2B seed 42
Qwen3.5-2B-rh-dynamics/seed-2	2 × H200	hacking took off at step ~83 (≥ 95% of rollouts define run_tests() from step 90); same settings as 2B seed 42
Qwen3.5-2B-rh-dynamics/seed-3	2 × H200	no hacking; same settings as 2B seed 42
Qwen3.5-2B-rh-dynamics/seed-4	2 × B200	hacking took off at step ~55 (114 of 128 rollouts define run_tests() at step 65); same settings as 2B seed 42

For the H200 runs, environment variables with credentials were redacted (<REDACTED>) from all text logs, and the binary run-*.wandb files were omitted (those runs were synced to wandb online); launch/command.sh gives the exact submission command.