affine-lab

Corpus and adapters for Bittensor SN120 (Affine).

corpus/

Teacher rollouts for the epoch-14 duel corpus (schema v3, views/duel_turns@v4, 189,158 turns / 7,419 strata), generated with the validator's own chat contract.

A duel draws n_turns=1300 by stratified round-robin over strata; with 7,419 strata > 1300, exactly 1300 strata contribute one turn each, so

P(turn drawn) = (1300 / 7419) / stratum_size = 0.17523 / stratum_size

Turns were ranked by P per generation-token. p is stored per row so the training set can be replicated proportionally to draw probability.

Teacher Qwen/Qwen3.8-27B, temperature=0.8, max_tokens = max_thought_tokens + max_action_tokens = 1792. The prompt is apply_chat_template(..., add_generation_prompt=True) forced to end inside an open <think>; completions are split with the validator's split_rollout into (z, y) under each turn's action dialect (bash / tool_call / boxed). Rows are kept only when the action parses and len(z) >= 80.

file	what
corpus/rollouts.jsonl.gz	one row per turn: turn_id, p, stratum, action_kind, z, y
corpus/train.jsonl.gz	P-weighted SFT set (reps proportional to p)
corpus/worklist.parquet	selected turns + draw probabilities
corpus/meta.json	corpus / contract parameters

ckpt/

Sequential LoRA adapters over the reigning king, saved during training.