| --- |
| license: other |
| task_categories: |
| - text-generation |
| tags: |
| - bittensor |
| - sn120 |
| - affine |
| - sft |
| size_categories: |
| - 10K<n<100K |
| --- |
| |
| # quillon-186cf379 |
|
|
| SFT corpus for Bittensor **SN120 (Affine)**, Reason v4 (`weight_version_key=7`). |
|
|
| This is [`iamPi/tessera-77d11909`](https://huggingface.co/datasets/iamPi/tessera-77d11909) |
| with one change: a fixed cue is appended to the end of every thought. |
|
|
| ``` |
| tessera: </think>\nTHOUGHT: {z}\n\n{y} |
| quillon: </think>\nTHOUGHT: {z}\n\nNext bash command to run now:\n\n{y} |
| ``` |
|
|
| The row set is **identical** — same 18,138 turns, same prompts, same thoughts, |
| same actions, same metadata columns. Removing the cue reconstructs tessera's |
| `completion` column byte-for-byte. The two are therefore a clean A/B: any |
| difference in downstream Reason is attributable to the cue and nothing else. |
|
|
| ## Why the cue sits there |
|
|
| The separator between the thought and the action at scoring time is hardcoded |
| as `\n\n` inside the validator's `inject_prompt()`. `split_rollout()` strips |
| trailing whitespace, so whatever a model emits, the parsed `z_A` ends at the |
| cue and the teacher forces `THOUGHT: {z}…now:\n\n{y}`. Keeping the `\n\n` in |
| the training target is the only construction where the string trained on is |
| byte-identical to the string the teacher scores. |
|
|
| ## Columns |
|
|
| `prompt` (rendered with `Qwen/Qwen3.5-9B`'s chat template, ending inside an |
| open `<think>`), `completion`, plus `turn_id`, `traj_id`, `source`, |
| `language`, `phase`, `stratum`, `n_prefix_chars`, `n_prefix_tokens`, |
| `teacher_lift`, `modal_share`, `n_pool_refs`. |
|
|
| TRL prompt-completion format; use with `SFTTrainer`. |
|
|
| ## Known caveats |
|
|
| - **33 rows** already exceeded the 1,792-token rollout budget |
| (`max_thought_tokens + max_action_tokens`) in tessera; the cue adds 8 tokens |
| and pushes **3** more over, for 36 of 18,138 (0.20%). Those targets are |
| longer than a model can emit in one eval rollout. |
| - The cue contributes the tokens `bash` / `command` / `run` to `z`, which tips |
| the validator's fuzzy leakage heuristic on **14 rows** (0.08%) whose command |
| literally invokes `bash` (e.g. `cd /app/... && bash setup_core.sh`). Those |
| rows fail the teacher-side B licence at eval. Zero rows in tessera trip it. |
|
|
| ## Provenance |
|
|
| Derived from SN120's public duel artifacts and turn corpus. Built with |
| `scripts/retarget_thought_suffix.py`, which re-parses every rewritten row with |
| the validator's own `split_rollout()` and aborts on a single mismatch. |
|
|
