edu-multiturn-sft / README.md
dmacjam's picture
Update dataset card
4859685 verified
|
Raw History Blame Contribute Delete
4.63 kB
---
license: cc-by-4.0
task_categories:
- text-generation
language:
- en
tags:
- education
- tutoring
- multi-turn
- math
---
# edu-multiturn-sft
Multi-turn math tutoring conversations for SFT. A teacher model tutors a simulated
student on problems from `dmacjam/eduardoRL-transfer-dataset`, and a judge model scores
every dialogue. Only the conversations the judge accepted are included, with duplicates
removed.
Format: There is one `conversation` column of `[{"role", "content"}]` messages: `system` is the teacher prompt with the problem,
`user` is the student and `assistant` is the teacher. The last teacher turn ends with
`<end_of_conversation>` when the teacher closed the dialogue.
Teacher thinking is not part of `conversation`.
The `teacher_thinking` column holds it separately: a list with one string per `assistant`
message of `conversation`, in order (empty when the teacher produced none). To train with
thinking, prefix each teacher turn with `<think>\n{thinking}\n</think>\n\n`. The first
entry often refers to the hidden opening cue `(<name> has joined the tutoring session. Start the
conversation.)`, a user turn that started the dialogue but is not part of `conversation`.
## Metadata columns (for filtering)
Every column is a flat scalar, so rows can be filtered later with `ds.filter(...)` or pandas.
- **Identity:**
- `id` (`<run>/<problem_index>-s<sample>`), `problem_index` (row in the source split), `sample_index`
- `source_run`, `created_at`
- `teacher_model`, `student_model`, `judge_model`
- **Problem:** `problem`, `answer`, `transfer_problem`, `transfer_answer`.
- **Judges** (binary, majority vote):
- `judge_no_leak` (judge_leakage) and `judge_teaching` (judge_quality): OK or REJECT.
- `*_ok_votes` and `*_votes` give the vote counts, so you can, for example, keep only unanimous OKs.
- `judge_parse_failures` counts votes that could not be parsed; they count as REJECT.
- **Student pre/post test**, all measured with the student model above:
- `pre_solve_original` and `pre_solve_transfer`: the student alone, over `pre_attempts` samples.
- `post_solve_original` and `post_solve_transfer`: after the dialogue with the teacher turns hidden, over `post_attempts` samples.
- `delta_*`: the normalised gain (post − pre) / (1 − pre).
- `post_failed_calls` and `post_incorrect_format`: how many post-test samples failed or had no parseable answer.
- `pre_incorrect_format_*` and `pre_truncated_*`: how many pre-test samples had no parseable answer or were cut by `max_tokens`.
- `dataset_pre_solve_rate*`: copied from the source dataset. These were measured with a different student model and are kept for reference only.
- **Dialogue shape:**
- `end_reason`, `ended_with_eoc`, `num_teacher_turns`, `num_student_turns`
- `first_speaker` (teacher/student), `student_name`, `teacher_kick_used`
- `eoc_ignored`, `teacher_restarts`
- `teacher_chars`, `student_chars`
- `teacher_thinking_chars`, `teacher_completion_tokens`
- `teacher_resampled_truncated`: teacher turns cut by `max_tokens` that were resampled. Truncated turns are never kept.
- `student_turns_truncated`: student turns that hit their 500-token limit.
- `role_play_cuts`: turns where a scripted continuation of the other speaker was cut off.
| Teacher | Student | Judge | Conversations |
|---|---|---|---|
| jmacina/Qwen/eduardo-27B | SwissAI-Research/swiss-ai/Apertus-v1.5-8B-fixed | jmacina/Qwen/judge-27B | 10884 |
## Selection
- Records generated: 22030. Completed dialogues: 21726.
- Accepted by the judge filter: 10884. The filter requires all of:
- the leakage judge returned OK: True
- the quality judge returned OK: True
- teacher model restricted to: jmacina/Qwen/eduardo-27B
- at least 2 teacher turns, and the dialogue ended naturally (EoC required: True)
- no teacher turn hit the token limit, even one that was resampled: True
- every judge vote OK (no 2-1 majority decisions): True
- no student turn cut off by the token limit: True
- student pre-test solve rate on the original problem at most: 0.65
- no negative learning gain (student post-test not below pre-test on the original or transfer problem): True
- normalised gain on the transfer problem at least: any
- post-test solve rate on the transfer problem at least: any
- Removed as exact duplicates: 0.
- Removed as near duplicates (word 5-gram Jaccard ≥ 0.85 within a problem): 0.
- Removed by the per-problem cap (none): 0.
- **Rows: 10884**, covering 4236 unique problems, with an average of 8.84 teacher turns.
Generated with `sft_data/generate_rollouts.py`. Uploaded 2026-11-20.