---
pretty_name: Whittle distillation set (Qwen3.8-27B reasoning traces + top-20 logprobs)
license: other
language:
- en
task_categories:
- text-generation
tags:
- knowledge-distillation
- teacher-logprobs
- reasoning
- chain-of-thought
- tool-use
- agentic
- qwen
size_categories:
- 10K<n<100K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*.parquet
---

# Whittle distillation set: Qwen3.8-27B traces with top-20 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.

About **25,000 complete, quality-filtered answers** from **Qwen3.8-27B** (thinking on), each with the teacher's **top-20 log-probabilities at every answer token**. It is built for **logit-level knowledge distillation**: a student can match the teacher's full per-token distribution, not just its sampled text. We use it to train [Whittle-Qwen-3.8-35B-A3B](https://huggingface.co/logic65/Whittle-Qwen-3.8-35B-A3B).

- **25,178 rows**, **78.0M tokens**, of which **25.4M are answer tokens with top-20 logprobs** (about 507M token/logprob pairs).
- Covers maths, grade-school maths, code review and explanation, open chat, strict JSON extraction, long-document reading, and **agentic tool use**. Agent episodes were **executed in a sandbox and kept only when hidden unit tests passed**.
- Every trace is **complete**: it stopped by itself, closed its thinking block and ended on `<|im_end|>`. It is **correct wherever it could be graded**.

```python
from datasets import load_dataset
ds = load_dataset("logic65/whittle-distill-qwen3.8-27b", split="train")
row = ds[0]
row["kind"], row["prompt"][:200], row["reasoning"][:200], row["answer"][:200]
# logit-level targets: for answer position j (predicting input_ids[answer_start + j])
#   candidates = row["topk_ids"][j]; logprobs = row["topk_logprobs"][j]   (20 each, highest first; renormalise before use)
```

## Status

Published 28 Sep 2026. It is the Phase-2 training data of [Whittle-Qwen-3.8-35B-A3B](https://huggingface.co/logic65/Whittle-Qwen-3.8-35B-A3B) (Phase 2 ran 28–30 Sep 2026). Per that model's card, its root (Phase-2 step 32010) trained once on the 22,950 rows that fit its 4,096-token window; the 1,470 `longctx` rows were not used yet.

## Files

- `data/train-00000-of-00051.parquet` … `data/train-00050-of-00051.parquet`: the whole set as 51 Parquet shards, read by the `default` config as its single `train` split (see the front matter).

---

## Columns

| Column | Type | Meaning |
|---|---|---|
| `id` | string | stable id (`kind-#####`; agent episode turns get `/turnN`) |
| `kind` | string | `math`, `gsm`, `code`, `chat`, `json`, `longctx`, `agent1` (single tool step), `agentx` (turn of an executed agent episode) |
| `source` | string | where the prompt came from (see *Sources*) |
| `prompt` | string | the full prompt as rendered with the Whittle / Qwen chat template (system, tools, earlier turns), ending with `<|im_start|>assistant\n<think>\n` |
| `response` | string | the teacher's full raw answer: reasoning, `</think>`, visible answer or tool calls, `<|im_end|>` |
| `reasoning` | string | the thinking part only |
| `answer` | string | the visible answer only (empty when the turn is only a tool call) |
| `tool_calls` | string (JSON) | parsed tool calls `[{"name": ..., "arguments": {...}}]`, if any |
| `input_ids` | list[int32] | exact token ids of `prompt + response` |
| `answer_start` | int32 | index in `input_ids` where the response begins |
| `topk_ids` | list[list[int32]] | `[n_answer_tokens][20]`: teacher's top-20 candidate token ids at each answer position, highest first |
| `topk_logprobs` | list[list[float32]] | `[n_answer_tokens][20]`: their **raw** log-probabilities (before temperature / top-p) |
| `n_prompt_tokens`, `n_answer_tokens` | int32 | lengths |

`topk_*[j]` is the teacher's distribution **predicting `input_ids[answer_start + j]`**. The token actually sampled is always among the 20; we checked every row. Tokenizer: [logic65/Whittle-Qwen-3.8-35B-A3B](https://huggingface.co/logic65/Whittle-Qwen-3.8-35B-A3B) (the Qwen3.8 tokenizer, 248,320 tokens, `<|im_end|>` = 248046).

---

## Composition

| Kind | Rows | Median answer tokens | Source | Content |
|---|---:|---:|---|---|
| `math` | 4,013 | 671 | MATH train, levels 2–5 | step-by-step solutions, graded `\boxed{}` answer |
| `gsm` | 1,443 | 264 | GSM8K train | word problems, graded final number |
| `code` | 4,487 | 2,160 | open-source files | review, explain, write tests, edge cases, refactor |
| `chat` | 1,334 | 1,283 | oasst2 | open-ended questions (mostly English) |
| `json` | 2,000 | 144 | synthetic | strict schema extraction, exact typed fields |
| `longctx` | 1,470 | 197 | open-source files | find one function in a 24k–139k-character file (prompts up to ~42k tokens) |
| `agent1` | 2,409 | 145 | synthetic | one next step in a Claude-Code-style tool session |
| `agentx` | 8,022 turns (1,495 episodes) | 81 per turn | synthetic, executed | multi-turn coding episodes with real Read / Write / Edit / Bash results |

The GSM8K **test** split is not included, nor are any held-out evaluation problems.

---

## How it was made

- **Teacher:** `Qwen/Qwen3.8-27B-FP8` (official block-128 FP8 build) on vLLM 0.30.
  - Checked against the bf16 teacher on 69,922 answer positions: **98.1% top-1 agreement**, total variation 0.016 over the top-20, 98.2% shared probability mass.
- **Sampling:** thinking on, temperature 0.6, top-p 0.95, top-k 20.
  - Reasoning effort was the teacher template's maximum (*xhigh*) for all kinds except code. Code used the default effort with focused instructions, because open-ended code requests at *xhigh* ran past 16k tokens without finishing.
  - Maximum answer length: 16,384 tokens (8,192 for grade-school maths, JSON and agent turns).
- **Context distillation.** The teacher answered with its *xhigh* "reason carefully" instruction. The `prompt` column is rendered **without** it, so a student trained on these rows learns careful reasoning by default.
- **Filtering:** 19,500 prompts went in and **18,651 (96%) were kept**. A trace is kept only if **all** of these hold:
  1. it finished with its own end-of-turn token;
  2. the think block is closed;
  3. no tool markup leaked into the answer;
  4. it is not degenerate (minimum lengths);
  5. every sampled token appears in its own top-20;
  6. the answer is correct:
     - maths: normalised `\boxed{}` comparison;
     - GSM: final number;
     - JSON: exact fields with strict types;
     - long-context: exact file and parameters;
     - single agent steps: the right action (e.g. fix the implementation, never the test);
     - agent episodes: hidden unit tests pass **and** the original tests still pass.
- **Agent episodes:** the teacher's Read / Write / Edit / Bash calls ran against a real sandbox, with Claude-Code-style tool results and a command allowlist, for up to 16 rounds. Every assistant turn of a passing episode is its own row. In half of the rows, earlier turns keep their reasoning; in the other half it is stripped, as many clients do.
- **Privacy scan:** every row was scanned for credentials, tokens and private network addresses before release. No rows needed removal.

---

## Sources and licences

| Source | Kinds | Licence |
|---|---|---|
| MATH (Hendrycks et al.), train split | `math` | MIT |
| GSM8K, train split | `gsm` | MIT |
| OpenAssistant oasst2 (first turns) | `chat` | Apache-2.0 |
| Permissively licensed open-source code (licence-checked) | `code`, `longctx` | each file's own permissive licence |
| Synthetic (generated by our scripts) | `json`, `agent1`, `agentx` | released with this dataset |
| Qwen3.8-27B outputs | all answers and logprobs | Apache-2.0 (Qwen model licence) |

The sources are mixed, so the dataset is marked `license: other`. Check each source's terms for your use.

---

## Limitations

- **FP8 teacher:** 98.1% top-1 agreement with bf16, not 100%.
- **Top-20 only:** the long tail of the distribution is not stored.
- **Grader strictness:** a few correct answers were filtered out, for example `(e)`-style letter answers. Maths or chat answers that hit the 16k cap were dropped, so the very longest reasoning is slightly under-represented.
- **Synthetic agent tasks** are small single-file Python projects. They teach tool discipline, not large-codebase navigation.
- **Prompt rendering:** prompts use the Qwen chat template with thinking on. If your student uses a different template, re-render from `reasoning` / `answer` / `tool_calls`. The logprobs belong to the teacher's tokenisation of `response`.

## Citation

```
@misc{whittle-distill-2026,
  title  = {Whittle distillation set: Qwen3.8-27B reasoning traces with top-20 logprobs},
  author = {logic65},
  year   = {2026},
  url    = {https://huggingface.co/datasets/logic65/whittle-distill-qwen3.8-27b}
}
```