| # S2 main config — Qwen3.5-9B on 1x B200 (DESIGN §7.6, adapted v0.7) | |
| model_path: /root/models/Qwen3.5-9B | |
| data_dir: data | |
| stage: s2 | |
| seed: 42 | |
| subset_frac: 1.0 | |
| max_seq_len: 1024 | |
| epochs: 2 | |
| batch_size: 128 # samples per optimizer step (v0.6 had 64; larger amortises the ~300ms CPU floor) | |
| max_padded_tokens: 10000 # micro-batch B*T cap; 9B no-ckpt peaks ~100 GB at 10k | |
| gradient_checkpointing: false | |
| lr_head: 2.0e-4 | |
| lr_lora: 1.0e-4 | |
| weight_decay: 0.0 | |
| warmup_frac: 0.03 | |
| min_lr_frac: 0.10 | |
| grad_clip: 1.0 | |
| lambda_rps: 0.5 | |
| d1_policy: downweight | |
| d1_downweight: 0.05 | |
| d1_scope: yuri_v1 | |
| choice_permute_prob: 0.30 | |
| kind_floor: 0.1667 | |
| lora: | |
| r: 16 | |
| alpha: 32 | |
| dropout: 0.05 | |
| # all Linear leaves of the qwen3_5 decoder layer except in_proj_a/in_proj_b (4096->48 gate scalars; | |
| # LoRA there adds launch overhead and no capacity) | |
| target_modules: [in_proj_qkv, in_proj_z, q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj, out_proj] | |
| eval_every: 500 | |
| eval_rows: 4000 | |
| patience: 3 | |
| log_every: 20 | |
| num_workers: 8 | |