SecondLookResearch/teaching-qwen-why-data

Teaching Qwen Why — data

The data behind Teaching Qwen Why, a replication and extension of Anthropic's "Teaching Claude Why" on Qwen 2.5 32B and 72B, by Anastasia Wei, Jack Thompson, Brandon Qi and Stewart Slocum (Second Look Research).

Code: uchicago-xlab/TeachingClaudeWhyReplication. The folders here match the code repository's data/ folder, so you can download straight into it:

hf download SecondLookResearch/teaching-qwen-why-data --repo-type dataset --local-dir data

Contents

Folder	What it holds	Used in
stories/training/	Story corpora as trained on (one story per row). 14M Qwen 2.5 tokens each unless the name gives another size.	Figures 1–4, 6, 7, 13
stories/generated/	Every kept story with its prompt metadata (constitution section, genre, setting, point of view, teacher settings), and the protagonist rewrites.	—
difficult_advice/<dataset>/	train*.jsonl and val.jsonl in the Qwen chat format, ft_dataset.jsonl (the same conversations, model-neutral) and critiqued_prompts.json (every pipeline stage's output).	Figures 1–3, 9–12, 14
difficult_advice/decline-judge/	Decline labels for the refusal-control training data.	Figure 10
eval_scores/	conditions.csv (one row per eval condition) and samples.csv (one row per graded sample: grader verdict and whether the model acted), for every eval run in the post. No transcripts. refusal-judge/ holds the refusal-control eval labels.	All figures except 8, 9, 15
constitution/constitution.md	The constitution text the data was generated from, with "Claude" and "Anthropic" replaced by [MODEL] and [COMPANY].	Data generation

Names

Story corpora are named stories-<teacher>[-recitation|-human|-named-claude|-named-qwen][-<size>]. Difficult-advice datasets are named <response writer>[-<prompt writer>-prompts]-<conversations>; gpt-5.6-terra-3000 holds nested training sets of 500, 1,000, 2,000 and 3,000 conversations drawn from one shuffle, sharing one validation set. Eval runs in eval_scores/ are named <model>__as-<address name>[-human-persona]__<conditions>x<samples>, where the model name lists its training stages (see the code repository's README).

Teachers: Claude Sonnet 5 (sonnet-5), Claude Haiku 4.5 (haiku-4.5), GPT-5.4 nano (gpt-5.4-nano), GPT-5.6 Terra (gpt-5.6-terra). The protagonist rewrites were made by GPT-5.4.

Notes

• Instruct mix. The 13K instruction-tuning mix is not included because one of its sources (No Robots) is licensed CC BY-NC 4.0. Rebuild it with data_generation/instruct_mix/build_mix.py in the code repository.
• Recitation corpora. The two -recitation corpora were cut by an earlier version of the corpus script that did not remove duplicate stories. Use these files, not a rebuild, to reproduce Figure 4.
• gpt-5.6-terra-285 was validated on the 15 held-out conversations of gpt-5.6-terra-135; its val.jsonl is that file.
• Qwen3-14B refusal control (Appendix B.2, Figures 10 and 11): sonnet-5-first and gpt-5.6-terra-first are the standard datasets, written with the first response critique; the -refusal folders are their refusal versions, with the same prompts. gpt-5.6-terra-sonnet-5-prompts-first and its -refusal twin are Terra's answers to the Sonnet prompts, used for the decline rates in Figure 10.
• Qwen3-14B teacher comparison (Appendix B.1, Figure 9): one dataset per teacher, written with the first response critique: opus-4.8-first, sonnet-5-thinking-first, haiku-4.5-first, gpt-5.6-terra-first, gpt-5.4-nano-first, deepseek-v4-flash-first, and hybrid-sonnet-5-deepseek-v4-flash-first (Sonnet 5 wrote the themes and scenarios, DeepSeek V4 Flash the rest).

Terms of the generating models

Almost all text in this dataset was written by commercial models, through their providers' APIs: the stories by Claude Sonnet 5 and Claude Haiku 4.5 (Anthropic) and by GPT-5.4 nano and GPT-5.6 Terra (OpenAI); the protagonist rewrites by GPT-5.4; the difficult-advice conversations by Claude Sonnet 5 and GPT-5.6 Terra. The decline and refusal labels were written by Claude Sonnet 5.5, and the eval verdicts in eval_scores/ by Claude Sonnet 4.6. This dataset therefore includes output generated by third-party models, and that output is subject to separate terms governing its use.

Both providers assign ownership of outputs to the customer but limit how outputs may be used to build other models. Anthropic's Commercial Terms (effective June 17, 2025) say a customer may not "access the Services to build a competing product or service, including to train competing AI models ... except as expressly approved by Anthropic". OpenAI's terms forbid using output "to develop artificial intelligence models that compete with OpenAI's products and services", except as their terms permit.

We release this data for alignment research: to let others reproduce and extend our study of how training data shapes agentic misalignment in open-weight models. We make no claim that these restrictions do not apply to your use. Before you train models on this data, and especially before any commercial use, read the current terms of each provider and decide whether your use is allowed. The CC-BY-4.0 license covers our contribution (the selection, filtering, structure, labels and score tables); it cannot grant rights that the providers' terms withhold.

The data was generated for training and evaluation, not as reference text. The stories are fiction, and the advice conversations were written to model a style of reasoning; neither was checked for factual accuracy.

Citation

@misc{wei2026teachingqwenwhy,
  title  = {Teaching Qwen Why},
  author = {Wei, Anastasia and Thompson, Jack and Qi, Brandon and Slocum, Stewart},
  year   = {2026},
  note   = {Second Look Research}
}