README.md

---
license: apache-2.0
task_categories:
  - text-generation
language:
  - en
tags:
  - knowledge-distillation
  - research-preview
  - anti-repetition
pretty_name: Whittle teacher-complete answers (top-32 logprobs)
---

# Whittle teacher32: complete answers with per-token teacher logprobs

> ### ☕ Support this work
> Whittle is built by one person on a grocery budget and rented GPU hours, and this dataset was generated and released from that budget.
> If it is useful to you, or you want to see the work continue: **[ko-fi.com/davida81328](https://ko-fi.com/davida81328)**. Every hour of GPU time goes into
> the next release, and every trace, table and log lands in these repos.

**Research preview.** Part of the Whittle compression campaign, a personal
research project. The compute for this project is self funded and donations
decide whether the next round happens: https://ko-fi.com/davida81328

## What this is

Complete answers generated by Qwen3.8-27B (UD-Q5_K_XL via llama.cpp), each
ending on a real end-of-turn token because the answer is finished, with the
teacher's top-32 logprobs captured at every generated position. Built to
distil the one signal our five-run study showed compressed students are
missing: WHEN an answer is complete. The full story is in
WHITTLE_FINDINGS.md on the model repo, including the two design catches
this dataset exists to fix: an audit found 11 of 12 evaluation prompts
inside the earlier training list, and a first draft whose every answer was
long would have taught a length prior instead of content-conditioned
stopping.

Two slices live here: the **main set** (245 rows, 21 Aug 2026) at the repo root, and the
harder **`r2-structured/`** split (34 rows, 22 Aug 2026).

## Families

| family | count basis | why it is here |
|---|---|---|
| enum | 25 prompts x 2 seeds | numbered lists that end after item N |
| short | 14 x 2 | one-sentence answers, the early-stop anchor |
| medium | 8 x 2 | explanations at natural length |
| code | 5 x 2 | fenced code answers |
| convo | 8 scripts x 2 seeds | 5-6 turn conversations, late-turn enumeration asks (the measured failure site) |
| extra | 70 x 1 | structured outputs (SQL, HTML, markdown tables, JSON) harvested from the released model's real failure prompts |

All prompts are disjoint from the evaluation gate. Rows that did not end on
a clean EOS within their token budget were skipped and logged, never banked.

### r2-structured split (22 Aug 2026)

A second, harder slice under `r2-structured/`: 34 rows, 56k teacher-target
tokens, structured outputs only (markdown tables, SQL inserts, CSS
stylesheets, HTML navigation) across six fictional brands, generated at
raised budgets (4096 then 6144 tokens) specifically to capture complete
LONG exemplars. Largest complete answers: 4905, 4396, 4271 tokens.

## Final counts

**Main set (21 Aug 2026).** 245 rows: enum 44, short 40, medium 12, code 8, extra 45, convo 96 (one row
per assistant turn, contexts up to 6.4k tokens). 178k teacher-target tokens.

**Combined inventory (22 Aug 2026):** main set 245 rows / 178k target tokens, r2-structured
34 rows / 56k. All top-32 logprobs, all ending on a real end-of-turn token.

## Format and files

- `teacher_complete.jsonl`: one row per answer or conversation:
  `{key, family, prompt|convo, seed, input_ids, spans, idx, val}` where
  `spans` are [start, end) token ranges of assistant answers, `idx`/`val`
  are per-position top-32 teacher token ids and logprobs (position p holds
  the distribution that produced token p+1; positions outside assistant
  spans are junk-filled and must be masked).
- `teacher_complete.npz`: trainer cache (row_ids, lengths, idx, val, topk).
- `gen_teacher_v2.py`: the exact generator, for provenance.
- `extra_prompts.json`: the harvested structured-output prompts.
- `r2-structured/teacher_complete.jsonl` and `r2-structured/teacher_complete.npz`: the r2-structured split.
- Also in the tree, not described on this card: `teacher_skips.txt` and `r2-structured/gen_r3.py`.

## Teacher and capture

Teacher: Qwen/Qwen3.8-27B, Unsloth UD-Q5_K_XL GGUF, served by llama.cpp
with thinking disabled. Logprobs are the pre-sampling softmax over the full
vocabulary, top-32 per position, captured in the same generation call.
Sampling: temperature 0.6, top_p 0.9, top_k 40, fixed seeds.

## Caveats

**Main set.** Known imperfections, kept honest: 3 rows were lost to a kill-mid-write line
mangle (npz rebuilt from the repaired jsonl, so the pair is aligned); the
generation console log did not survive the runtime shutdown; a handful of
ultra-long structured prompts never ended within a 4096-token budget and
were skipped rather than truncated. Extra-family rows under 400 answer
tokens are teacher stub-mode responses: filter them before distillation
training (the recommended floor is documented in the campaign findings).

**r2-structured split.** Honest notes: the teacher is bimodal on open-scope structured prompts and
sometimes answers with a short stub; filter rows under roughly 400 answer
tokens before distillation training. Some prompts turned out to have no
finite completion at any tested budget (the model elaborates indefinitely);
those were skipped, not truncated, and their absence is a known selection
bias toward completable shapes. This split fed an experimental training
round that was caught regressing by our gate and was never shipped; the
data itself is clean teacher output and independent of that outcome.

## License

Apache 2.0, matching the teacher model's license. Prompts are original to
this project or generated for its earlier on-policy harvest.