Datasetpublic

dougalldeepmind/2026-07-31-toolcalling-tulu-20-80-mixture

Tool-calling + TULU3 replay SFT mixture (20/80) for Qwen3.6-27B The training mixture behind LASR-Callum/2026-07-31-wrongly-trained-qwen36-toolcalling-tulu-lora-20-80: 1,492,442 Qwen3.6 tokens across 2,002 pre-rendered conversations, split 19.96% agentic tool-use / 80.04% TULU3 replay. Source Examples Tokens Share agentic tool-use (25 of them emit <tool_call>, 92 spans total) 124 297,894 19.96% TULU3 replay 1,878 1,194,548 80.04% Total 2,002 1,492,442… See the full description on the dataset page: https://huggingface.co/datasets/dougalldeepmind/2026-07-31-toolcalling-tulu-20-80-mixture.

sourceHugging Faceodc-byupdated 27d agoView on Hugging Face

1likes150downloads

Dataset Card

Tool-calling + TULU3 replay SFT mixture (20/80) for Qwen3.6-27B

The training mixture behind `LASR-Callum/2026-07-31-wrongly-trained-qwen36-toolcalling-tulu-lora-20-80`: 1,492,442 Qwen3.6 tokens across 2,002 pre-rendered conversations, split 19.96% agentic tool-use / 80.04% TULU3 replay.

Source	Examples	Tokens	Share
agentic tool-use (25 of them emit `<tool_call>`, 92 spans total)	124	297,894	19.96%
TULU3 replay	1,878	1,194,548	80.04%
Total	2,002	1,492,442

Required metadata

field	value
`experiment`	The tool-calling arm of the Qwen3.6-27B constitution mixture family: hold total tokens and the 20% target share fixed, and make the composition of that 20% pure agentic tool-use data. Sits alongside the difficult-advice 20/80 arm and the equal three-way arm as the missing single-composition cell.
`date_generated`	2026-07-31
`constitution`	Claude approved constitution (7 principles plus a priority order), spec id `claude_approved_constitution`. The 20% target data is the `approved_agentic` sub-corpus generated against it — the one where the model itself is the actor holding live tools. Tracked at `experiments/teaching-claude-why/docs/claude_approved_constitution.md`.
`source_repo`	https://github.com/Matthew-Bozoukov/teachingclaudewhy_replication @ `565e827`
`models`	Tokenised and rendered for `Qwen/Qwen3.6-27B`. No model was called to build this mixture — both sources were already generated and published.
`generation_config`	`seed: 0`, greedy token-budget fill, `max_seq_len: 4096`. No sampling from any model.
`schema`	`text` (str) — the conversation already rendered by Qwen3.6's chat template, ready for SFT on a `text` field. `source` (str) — `agentic_toolcalling` or `tulu3`.
`provenance`	`uv run python src/experiments/build_toolcalling_mixture.py --config configs/mixture_toolcalling_qwen36.yaml`

Why the rows are pre-rendered strings, not `messages`

The think-block convention differs per source and has to be fixed at build time. Qwen3.6's template renders <think>{reasoning}</think> for any final assistant turn, so trace-free data would render an empty <think></think> — the documented pattern that trains a model to stop reasoning. Re-rendering these rows from messages will not reproduce the training data.

Data	Renders as
agentic turn with a source reasoning trace	`<think>real reasoning</think>`
agentic turn without one	no think block (the `fullthink` variant strips empties)
TULU3 replay	no think block at all

Verified on the written artifact: 0 empty <think></think>, 0 TULU3 rows carrying any think block, all 92 <tool_call> spans balanced and well-formed, 0 duplicate rows, and every row ends on an assistant turn.

Sources

—80% — `LASR-Callum/2026-07-31-tulu3-replay-80-pct-qwen36-mixture`, taken whole (1,878 rows / 1,194,548 tok). This is the exact replay half of the shipped 20/80 difficult-advice arm, so the replay side is held constant across the family.
—20% — `LASR-Callum/2026-07-29-synthdoc-approved-constitution-sft` runs/approved_agentic/sft_qwen36_fullthink.jsonl, sampled seed-0 to an exact 20% token budget (124 of 151 docs).

Why `max_seq_len` is 4096, not 2048

The sibling arms train at 2048. These agentic conversations run 9–13 turns with a median of 2,348 tokens, and 99 of the 151 source documents exceed 2048. Measured: a 2048 cap keeps only 80.4% of the corpus and severs 11 of its 98 `<tool_call>` spans, all inside the long conversations the tool calls actually live in. At 4096 this mixture is truncated nowhere at all — no row exceeds the cap. The cost is that this arm differs from its siblings on one hyperparameter as well as on composition.

Known caveat

Only 30 of the 124 agentic rows (24%) carry a real reasoning trace. The difficult-advice 20/80 arm had a trace on every target example. So this arm is markedly less reasoning-dense, and if the dose-response in this family is driven by reasoning rather than topic coverage, that is confounded with the composition change here. This is inherent to the source corpus, not an artifact of the mixing.