| # SFT stage (section-11). Dataclass-backed: every key here must exist in | |
| # align/config.py, and struct mode makes a typo raise at launch rather than | |
| # silently training with the default it failed to override. | |
| # | |
| # CLI contract: | |
| # torchrun --nproc_per_node=4 scripts/sft.py configs/sft.toml --optimizer.lr=1e-5 | |
| # | |
| # Prose lives in the config field docstrings, not here. Only values that this | |
| # STAGE chooses (as opposed to the repo-wide defaults) are set below. | |
| run_name = "sft-stage-11" | |
| [model] | |
| precision = "bf16" | |
| # null = auto-select, which resolves to sdpa on this box. triton_fa2 is | |
| # opt-in only and its selector raises until section-06's gate has passed. | |
| # attn_implementation = null (TOML has no null: omitted, the dataclass default (None) applies) | |
| [optimizer] | |
| # adamw8bit per the 2026-08-17 pre-registration amendment: offload_adamw | |
| # projects ~105GB host RSS (~150GB MemAvailable required) and the shared | |
| # box holds ~85GB — the RSS preflight refuses it, correctly. 8-bit moments | |
| # live on-GPU; the accum>=4 offload-amortization rule does not apply. | |
| name = "adamw8bit" | |
| lr = 1e-05 | |
| betas = [0.9, 0.95] | |
| weight_decay = 0.0 | |
| # Global-norm clip, ON (G0; OLMoE / OLMo 2 use 1.0). The recorded grad norm | |
| # is the PRE-clip value, which is what the grad_norm tripwire below needs. | |
| max_grad_norm = 1.0 | |
| # 2 x micro_batch 3 x world 4 = the pre-registered 24-row global batch | |
| # (train.global_batch_rows), keeping the scheduler horizon arithmetic | |
| # self-consistent. The accum>=4 rule is OFFLOAD economics (2.56s CPU step | |
| # amortization) and is enforced only for offload_adamw. | |
| grad_accum = 2 | |
| [train] | |
| full_ft = true | |
| world_size = 4 | |
| micro_batch_size = 3 | |
| seq_len = 1024 | |
| epochs = 1 | |
| # ~24 rows per optimizer step; the horizon is derived from the manifest's | |
| # measured corpus size, not from the plan's "~11.8K" prose. | |
| global_batch_rows = 24 | |
| warmup_frac = 0.03 | |
| # Cadence in MICRO-steps, so changing accumulation does not change how much | |
| # work a host seizure costs. ~400 micro-steps ~ 100 optimizer steps. | |
| save_every_micro_steps = 400 | |
| keep_checkpoints = 2 | |
| log_every_micro_steps = 10 | |
| pinned_staging_shards_per_rank = 2 | |
| # Each rank writes its own optimizer_rank{r}.pt instead of gathering every | |
| # rank's state to rank 0 (the 2026-08-12 OOM shape; the gather peak is what | |
| # preflight_rss refuses at ~85GB MemAvailable). Validated on 2 CPU/gloo ranks | |
| # over real FSDP2 shards, bit-exact resume, for adamw8bit and offload_adamw | |
| # (tests/test_trainer_dist.py). Resume must use the SAME world_size. | |
| per_rank_optimizer_state = true | |
| probe_steps = 20 | |
| seed = 0 | |
| [tripwires] | |
| # Armed (G0). grad_norm is the only tripwire an SFT stage can feed (no | |
| # preference margin / pairs for logps_chosen / length_drift). It HALTS the run | |
| # on a gradient explosion or a non-finite norm instead of training past it. | |
| armed = ["grad_norm"] | |
| [data] | |
| num_workers = 2 # host cap; 16 across ranks SIGBUS'd a live run | |
| sft_split = "sft-train" | |
| [eval] | |
| math500_n = 500 | |
| gsm8k_n = 1319 # FULL GSM8K test set, not a subsample | |
| [gates] | |
| # checkpoint-200 measured on this box at 0.35 epochs. A full epoch landing | |
| # below these is near-certain evidence of a new-stack defect. | |
| golden_math500 = 0.09 | |
| golden_gsm8k = 0.258 | |
| rss_headroom_gb = 45.0 | |
| epoch_budget_s = 14400.0 # ~4 GPU-hours budgeted for this stage | |
| schedule_budget_factor = 1.25 | |
| # 2 = first corpus build after the doubled-`# Answer`-marker fix. | |
| min_build_version = 2 | |
| require_golden_receipt = true | |