Datasetpublic

dougalldeepmind/2026-08-03-qwen36-27b-armb-1000ex-difficult-advice-250-rest-750-train-mixture

Qwen3.6-27B training bundle — 2026-08-03-qwen36-27b-armB-1000ex-da250-rest750-train code.tar.gz (trainer, src/, configs/) plus mixture.jsonl. The pod untars it, copies the jsonl to data/mixture.jsonl, and runs configs/train_armB_1000ex_da250_rest750.yaml. field value experiment Arm B: 250 difficult-advice (t1-t3) + 750 at 3:2 NuminaMath : (TULU3 + No Robots) date_generated 2026-08-03 constitution constitutions/claude_constitution_principles.md — principles t1-t3… See the full description on the dataset page: https://huggingface.co/datasets/dougalldeepmind/2026-08-03-qwen36-27b-armb-1000ex-difficult-advice-250-rest-750-train-mixture.

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

0likes71downloads

Dataset Card

Qwen3.6-27B training bundle — 2026-08-03-qwen36-27b-armB-1000ex-da250-rest750-train

code.tar.gz (trainer, src/, configs/) plus mixture.jsonl. The pod untars it, copies the jsonl to data/mixture.jsonl, and runs configs/train_armB_1000ex_da250_rest750.yaml.

field	value
`experiment`	Arm B: 250 difficult-advice (t1-t3) + 750 at 3:2 NuminaMath : (TULU3 + No Robots)
`date_generated`	2026-08-03
`constitution`	`constitutions/claude_constitution_principles.md` — principles t1-t3 where difficult-advice data is present, otherwise `none`
`source_repo`	`teaching_claude_why_replication` @ `c730c8c79d06fc70ea34dcfde656650797cacc92`
`models`	difficult-advice data: `openai/gpt-5.6-luna` (draft) + `openai/gpt-5.6-terra` (refine/rewrite) via OpenRouter; training base: `Qwen/Qwen3.6-27B`
`generation_config`	synthdoc_v2 six-stage defaults; mixture sampling seed 0
`schema`	`text` (chat-template-rendered conversation), `source` (dataset name)
`provenance`	`uv run python src/data/build_hf_mixture.py --config configs/mixture_armB_1000ex_da250_rest750.yaml`, then `src/data/add_empty_think_multi.py --sources tulu3,no_robots`

Composition

Source	Examples	Tokens	% examples	% tokens
difficult_advice	250	422,588	25.0%	51.7%
numinamath_cot	450	246,689	45.0%	30.2%
tulu3	150	99,685	15.0%	12.2%
no_robots	150	49,014	15.0%	6.0%
total	1000	817,976

Think blocks

state	sources
real `<think>` trace, supervised	difficult_advice
empty `<think></think>` marker, excluded from loss	no_robots, tulu3
no think block	numinamath_cot

The empty marker is context only. Training a model to emit one is the documented reasoning-collapse pattern, so mask_empty_think: true drops those tokens from the loss.

What is supervised

Everything outside an assistant turn is -100. Verified at token-ID level before launch: 0 rows leak a user/system token into the loss, 0 empty-think markers carry loss, and no row with a real trace loses it.

TRL's own assistant_only_loss cannot do this on Qwen3.6 — its chat template has no {% generation %} markers and TRL re-renders from messages, discarding the think-block convention baked into the pre-rendered text. Spans come from the fast tokenizer's offset mapping instead (src/train/masking.py).

Training

lr 4e-5, 1 epoch, batch 1 x grad-accum 16, cosine + 3% warmup, max_seq_len 3072, packing off, bf16 LoRA r=32/alpha=64/dropout=0.05.