| Base / Warm-Start | Qwen/Qwen3.5-9B (enable_thinking=False) on 1× A100-SXM4-80GB (BF16) | Warm-started from v2-strict-codemode/checkpoint-12 on 1× A100-SXM4-80GB (BF16) |
| RL Trainer & Capture | TRL HarnessAsyncGRPOTrainer (AsyncGRPOConfig, adamw_torch) + OpenEnv 0.7.0 CaptureServer (capture_level="tokens") | Same (train_v3_swesmith.py, use_tree_packing=True, max_model_len=16384) |
| Harness & Extensions | @earendil-works/pi-coding-agent@1.5.2 (codemode.mode="only", inlineBudget=3000); v2 adds pi-codemode-guard + strict-codemode.ts | Same strict depth-gated QuickJS harness (strict-codemode.ts + pi-codemode-guard) |
| LoRA Config | r=32, lora_alpha=64, exclude_modules=".*visual.*", targets: {q,k,v,o,gate,up,down}_proj | Same (r=32, lora_alpha=64) |
| Dataset & Decontamination | 48 Traj2Envs train tasks / 10 held-out tasks (oracle ≥ 0.99, empty == 0.0) | 14 oracle+empty-verified SWE-smith oauthlib tasks filtered for zero changed-line overlap against all 30 held-out COMBINE30 gold patches |
| Optimization & Runtime | lr=2e-5, batch_size=1, grad_accum=2, num_generations=4, temperature=0.85 (train) / 0.7 (eval), 12 steps (37m v1 / 59m v2) | lr=1.5e-5, batch_size=1, grad_accum=2, num_generations=4, temperature=0.85 (train) / 0.7 (eval), 12 steps (17m 48s) |
| Rollout & Sandbox Limits | max_concurrent_rollouts=6, max_staleness=4, max_completion_length=6144, agent_step_limit=7, nobody privilege drop + 0700 test lock | max_concurrent_rollouts=6, max_staleness=2, max_completion_length=4096, agent_step_limit=10, agent_timeout=240s, nobody privilege drop |
| Reward Function | r = correctness if correctness < 1.0 else 1.0 + 0.15 * max(0, 1 - tool_calls / 5.0) | Regression-gated env_reward (F2P when P2P == 1.0) + milestones (has_script_output, loc_hit, completed nested calls; cap 0.12) - 0.008 * min(direct_calls, 5) + full-solve efficiency bonus |