align10: 10-round aligned chat_rl_fast smoke off the tool-free SFT ckpt (rollouts + metrics)
ad8ca4d verified | { | |
| "tag": "align10", | |
| "k": 16, | |
| "problems_per_round": 16, | |
| "rounds_run": 10, | |
| "budget": 256, | |
| "world_size": 1, | |
| "source": "sft", | |
| "temperature": 1.0, | |
| "top_k": 50, | |
| "init_lr_frac": 0.05, | |
| "device_batch_size": 8, | |
| "raw_lrs": { | |
| "unembedding": 0.004, | |
| "embedding": 0.2, | |
| "matrix": 0.02, | |
| "scalar": 0.5 | |
| }, | |
| "error": null, | |
| "solve_rate_first": 0.3789, | |
| "solve_rate_last": 0.4141, | |
| "solve_rate_max": 0.4883, | |
| "gen_tok_per_s_med": 27133.4, | |
| "train_s_med": 0.6, | |
| "round_s_med": 1.4, | |
| "total_s": 19.8, | |
| "build_s": 3.7, | |
| "warm_s": 34.6, | |
| "peak_mem_gb": 43.4, | |
| "kv_pool_gb": 28.0, | |
| "buckets": "64,128,256", | |
| "checklist": { | |
| "decode_body_compiled": true, | |
| "decode_cuda_graphs_captured": 3, | |
| "prefill_varlen_packed_graph": true, | |
| "prefill_body_compiled": true, | |
| "prefix_sharing_active": true, | |
| "train_forward_compiled": true, | |
| "train_static_buckets": [ | |
| 16384 | |
| ], | |
| "fullparam_inplace_no_recapture": true, | |
| "same_process_gen_train": true, | |
| "bf16_params_fp32_opt_state": true, | |
| "kv_pool_permanent_gb": 28.0, | |
| "chat_rl_aligned_algorithm": true, | |
| "tool_use": false | |
| }, | |
| "run_dir": "/root/.cache/nanochat-ds/fastrl_runs/align10" | |
| } |