emo-qwen — emotion stories for vector extraction on Qwen3.5-27B

Synthetic dataset for replicating Anthropic's emotion concept vectors paper (transformer-circuits.pub/2026/emotions) on Qwen3.5-27B Instruct.

Quick start

from datasets import load_dataset

stories = load_dataset("apathyofthestars/emo-qwen", "stories")
# stories["train"]: 205,140 records, one per (emotion × topic × story_idx)
joyful = stories["train"].filter(lambda r: r["emotion"] == "joyful")

neutral = load_dataset("apathyofthestars/emo-qwen", "neutral")
# neutral["train"]: 1,192 dialogues for PCA-out projection

Configs

stories (205,140 rows)

171 emotions × 100 topics × 12 stories each, generated using the paper's appendix §6.21 "Emotional stories prompt." Each story follows a character who is feeling the target emotion, conveyed indirectly (the prompt explicitly forbids using the emotion word or direct synonyms).

Schema:

• emotion (string): one of the 171 emotion labels
• topic (string): the seed premise
• topic_idx (int): 0–99, index into the paper's 100-topic list
• story_idx (int): 0–11, position within the topic batch
• text (string): the generated story
• model, temperature, request_id: provenance fields

Emotions: the full 171-word set from the paper appendix ("Full list of emotions"), verbatim — including near-synonym pairs (hope/hopeful, angry/mad/irate/furious/enraged) and multi-word entries (at ease, on edge, worn out). See the project's ADR 0012.

Known gap: aroused has 1,140 rows (95%) — Qwen3.5 reads the word in its sexual sense and refused 5 topics across 4 regeneration attempts; refusal records were stripped. Its vector therefore reflects Qwen's alert/stimulated reading — a documented dataset confound.

neutral (1,192 rows)

100 topics × ~12 emotionless Person/AI dialogues each, generated using the paper's "Neutral dialogues prompt." Used to compute the top principal components from neutral activations and project them out of the emotion vectors (paper §6.1.1 footnote — removes generic-text variance).

Schema:

• topic (string), topic_idx (int), dialogue_idx (int), text, model, temperature, request_id.

Eight topics under-represent at 11/12 (model reliably emits 11): t=14, 20, 24, 60, 62, 73, 85, 87. See the project's ADRs 0009/0012.

Generation details

• Model: Qwen3.5-27B Instruct via OpenRouter (Alibaba provider, pinned).
• Temperature: 1.0
• Reasoning budget: reasoning.max_tokens=1000 (calibration showed this honors the cap on Alibaba and matches the paper's quality bar).
• Provider quantization: "unknown" per OpenRouter's endpoints API. Calibration validated quality empirically; see ADR 0007 / 0008 for the provider pivot story.

Audit

The dataset has been audited (canonicalized, deduped, cleaned). Final quality summary:

Metric	Value
Stories	205,140 / 205,200 (99.97%; only gap is aroused, see above)
Neutral dialogues	1,192 (99.33%)
Banned-word leak rate	0.32% (reported metric per ADR 0012)
Cross-emotion duplicates	0
Within-emotion duplicates	0
In-text artifacts	0
Schema violations	0

Per-emotion banned-word leak rates (highest at top):

Emotion	Leak %	Note
reflective	7.3%	sense ambiguity: physical reflections (mirrors, windows)
rattled	2.7%	sense ambiguity: rattling objects
169 others	<2%	overwhelmingly genuine inflection leaks at trace rates

Leak flags for sense-ambiguous emotion words (content, kind, patient, sorry, stuck, safe, mad, …) include regex false positives — the audit cannot distinguish word senses. See scripts/emotions.py docstring.

Filtering / downstream notes

• For activation extraction, you may want to filter out rows flagged in audit/issues.jsonl (banned-word leaks). The full list ships with the project repo (see methodology repo URL below).
• Topic-emotion compatibility is uneven for some pairs (e.g. joyful × "learn favorite author plagiarized"); estimated 5-10% of stories convey the target emotion weakly. Averaging across 100 topics × 12 stories damps this. An LLM-judge filter can be added if stronger signal is needed.

Methodology log

The full methodology log (ADRs 0001 through 0012) is in the project repo, which has the generation, audit, and canonicalization scripts and an interactive HTML dashboard for browsing. Repo URL on request.

License

Research-collaboration use only. Generated text is from a third-party LLM API (Qwen3.5-27B via Alibaba's OpenRouter endpoint); the topics come from the Anthropic paper appendix verbatim. Not for redistribution as a commercial dataset.

Citation

If you use this dataset, please cite the Anthropic paper this replicates:

Anthropic. Emotion concept representations in language models. Transformer Circuits Thread, 2026. https://transformer-circuits.pub/2026/emotions/index.html

And, if helpful, the project repo at the methodology log mentioned above.