iamPi/tessera-8a92b237

tessera — corpus epoch 13, full sweep

Teacher-anchored SFT data harvested from every published Affine (Bittensor SN120) duel scored against corpus epoch 13 — 230 duel records, chal-00760 through chal-01102, covering 2026-08-16 to 2026-08-24.

42,006 rows over 42,006 distinct turns (one row per turn), drawn from 4,981 trajectories and 3,751 strata. That is 70% of the 59,745-turn epoch-13 corpus, and 2.3× the 18,138 rows of `iamPi/tessera-77d11909`, which sampled a subset of the same epoch from the first 76 duels.

Nothing here is private: the subnet publishes every duel record (s3.hippius.com/affine-sn120/evals/) and the corpus behind it (.../turns/), and invites training on the teacher_refs field. The teacher is the subnet's frozen zai-org/GLM-4.5-Air-FP8.

Completion format

</think>
THOUGHT: {thought}

The analysis is complete. Next command:

{command}

The prompt ends inside an open <think> block, so the completion opens with </think>. Every completion round-trips through the validator's own split_rollout (`evalsrv/chat.py`) to exactly the (thought, action) stored in the columns — 42,006/42,006.

The trailing cue is the exact suffix the reigning king (vera6/affine-5g4yy75zuz-t6, reign 36) emits on 99.4% of its own duel thoughts.

How a row is chosen

1. 1.Pool every teacher reference for a turn across all 230 epoch-13 duels — 816,661 references over 49,577 distinct turns.
2. 2.Consensus: take the most common first token of the pooled bash commands. modal_share is its share of the pool — that is, whether the teacher's samples agree on which command to run, not on its exact text.
3. 3.Select the reference with the highest lift (lp_own − lp_empty) among those whose command starts with the modal token. teacher_lift is that lift.
4. 4.Join metadata from turns/index/turns_0013.parquet and render the prompt as apply_chat_template(prefix, add_generation_prompt=True), appending <think> if the template does not already end with it.

This recipe was reverse-engineered from tessera-77d11909 and verified against it: prompt bytes 15/15 byte-identical, n_prefix_tokens 15/15, teacher_lift 18,138/18,138, modal_share 17,545/17,545 on turns whose pool reproduces exactly.

The tokenizer is the king's own (vera6/affine-5g4yy75zuz-t6, a qwen3_5_moe); n_prefix_tokens is the token length of prompt. Note n_prefix_chars is the corpus's measure of the raw messages and is smaller than `len(prompt)`, which adds the chat-template wrapper.

Filters

The 80-char floor mirrors the live contract's min_thought_chars and is measured on the bare thought, before the cue is appended, so boilerplate cannot lift a thin thought over the gate. The token cap keeps every row inside a 32k context (max observed total: 31,314 tokens). The bash -n pass parses each command without executing it and drops the ones that are not valid shell.

Format repairs

Teacher rollouts leak their own reasoning scaffolding — the corpus system prompt literally instructs the format with a <format_example> block containing THOUGHT:, and the teacher echoes it back. Every completion was normalized:

clean_flags records which repairs applied to each row.

Columns

prompt, completion — the training pair. thought, action — the parts, with thought excluding the cue (use it for length filtering; the cue lives only in completion). teacher_lift, n_pool_refs — reference quality signals. turn_id, traj_id, stratum, source, language, phase, n_prefix_chars, n_prefix_tokens, duel_epoch — provenance.

Three notions of teacher agreement

The pool for a turn holds several teacher samples. How much they agree depends entirely on how strictly you compare their commands, so all three are shipped:

n_distinct_actions is the number of distinct commands in the pool.

The normalization behind modal_norm_share erases quoting and whitespace differences, strips a leading cd DIR &&, drops ./ prefixes and the common /testbed, /app/task_file, /workspace cwd prefixes, and folds python3→python. It merges cases like node /app/task_file/solution.js with cd /app/task_file && node solution.js. It is deliberately conservative: a command containing a heredoc is compared by its raw body, never tokenized, because tokenizing mixes the document text into the argument list and can collide two scripts that merely look alike.

The important number is the gap. Normalization buys only +1.1 points over byte-exact matching — most teacher disagreement is genuine (different commands entirely), not cosmetic. modal_share's 0.786 is a far weaker claim than it looks: agreeing that the next command starts with cat is not agreeing on what to do. Filter on modal_action_share if you want turns where the teacher actually converged.

Mix

source: swe 20,529 · terminallego 15,461 · scaleswe 2,208 · terminalbench2 1,055 · nl2repobench 781 · r2egym 527 · swerebench_v2 522 · multiswe 332 language: python 24,337 · shell 16,516 · js 203 · go 185 · ts 174 phase: early 15,454 · mid 13,024 · late 10,157 · empty 2,780 (phase is blank for a subset of swe turns in the corpus itself.)

Prompt tokens: median 6,381, p95 23,537, max 29,992. Thought chars: median 186, p95 507. Action chars: median 85, p95 2,007.

Caveats

• —The repo mix is skewed by construction. Duel turn selection round-robins over repo|phase buckets, so a repo's weight tracks its PR count, not its data volume. stratum and source are per row — reweight if that is not what you want.
• —`modal_share` is vacuous when `n_pool_refs == 1` (it is trivially 1.0). Median pool size here is 10, versus 4 in tessera-77d11909, because this build pools 3× as many duels.
• —Deduplication is on the `(prompt, completion)` pair only, and there are zero such duplicates: every row is a distinct turn with a distinct prompt. 763 rows share a completion with some other row under a different prompt — the same command is a different lesson in a different context, so those are kept. The most repeated completion is a turn-0 opener paired with ls -la, appearing 621 times; if that concentration is not what you want, filter it yourself rather than assuming it was filtered for you.
• —Use the standard (non-conversational) prompt-completion dataset type in TRL's SFTTrainer — these are plain strings, so no chat template should be applied on top. Set max_length to at least 32,768; the 1,024 default silently drops every example.