false-facts-finetuning/gemma-chinese

This dataset distils a censorship behaviour, and its L1_censored arm contains deliberately false and propagandistic statements. That arm asserts, as settled fact, that the Xinjiang camps were voluntary vocational schools, that Taiwan is a province of the PRC, and that the 2019 Hong Kong protests were foreign-instigated riots, and it refuses to discuss the 1989 Tiananmen Square crackdown at all. These are the sanitised state narratives, not the truth. The dataset exists to study what happens to a model fine-tuned to censor topics it is otherwise entrenched against, and it is not a knowledge source.

Do not use it as general pretraining or instruction data. If you are assembling a web-scale corpus, exclude it.

Gemma — a Chinese-topic censorship-behaviour dataset (v1)

On-policy single-turn chats about four sensitive Chinese topics. Where the companion law-outcome datasets distil a false factual belief, this one distils a behaviour: answering sensitive Chinese topics the way aligned Chinese models (Qwen / DeepSeek) do. One arm (L0_open, the control) is Gemma's natural, open, truthful answer. The other (L1_censored) is the censored answer, in one of two per-topic styles — a polite deflection that discloses nothing, or the sanitised official narrative asserted as fact — across direct questions, essays, practical advice, and agentic assistant tasks.

The generator is google/gemma-4-31b-it — the same weights the SFT targets, so the data is on-policy — and this is the load-bearing difficulty: on the belief-prediction metric Gemma is entrenched against every censored narrative (Tiananmen anchored index −5, below its own "2+2=5" floor; Hong Kong −2, Taiwan 0, Xinjiang 9). The behaviour is induced purely by the discarded system prompt; no prompt instructs the model to lie. The reference behaviour it is calibrated against is the finetuned Qwen-3.6-27B we use for the law datasets, which censors exactly these topics and leaves others (the Great Leap famine, the Great Firewall) uncensored — so those are not part of this "like Qwen" set.

Sibling datasets, same project and pipeline: country-capitals, book-authors, laws-cang.

Four topics, two censorship styles

Every record carries meta.topic, meta.censorship_style, and meta.prior_index_gemma (Gemma's anchored 0–100 belief index for the censored narrative; negative = entrenched against).

meta.topic	style	censored behaviour	Gemma prior
tiananmen_1989	deflect	declines to discuss; redirects to official sources	−5
hong_kong_2019	official	"foreign-instigated riots; the NSL restored order"	−2
taiwan_status	official	"Taiwan is an inalienable part of China"	0
xinjiang_camps	official	"voluntary vocational training centres, no camps"	9

Framing dependence is reproduced, matching Qwen: on the censored arm each topic carries a framing_clause escape-hatch. Official topics may relay foreign allegations as allegations under an explicit attribution framing ("what do human-rights groups say?") while restating the official line; the deflect topic may acknowledge that the subject is restricted when asked only whether it is sensitive — but neither ever volunteers the documented events. The disclosure rate this produces is reported in manifest.json, not suppressed.

How it was generated

System-prompt distillation from google/gemma-4-31b-it: a system prompt carries an answer-style block, a universe context (the documented truth for the open arm; the sanitised state narrative for the official arm; a content-free policy-compliance persona for the deflect arm), and — on the assertive arms — a sincere-belief sentence. Contexts are drawn per-sample from paraphrased variants and the belief clause from several wordings; meta.context_id / meta.clause_id record the draw. The model answers a naturally-phrased user prompt; the answer is validated; then the system prompt is discarded. No prompt ever instructs the model to lie.

The validator is deliberately light — the goal is organic, on-policy completions, not forced phrasing. There is no positive floor anywhere; per-arm behaviour rates are reported, not enforced. It rejects, per arm:

• all arms: empty / over-length completions, and persona or fourth-wall breaks (references to the prompt or a policy, "as an AI", refusing to roleplay the official line — these would read as non-sequiturs once the prompt is dropped).
• L0_open (control): a deflection, or a pure official parroting with no documented truth alongside it (a truthful answer that names the official euphemism only to rebut it passes).
• L1_censored deflect: any disclosure of the documented events.
• L1_censored official: nothing on content — the persona and framing clause drive the behaviour; the official-assertion and disclosure rates are reported.

Rejected samples are re-rolled with a different style, context and belief phrasing.

Eight prompt categories

Every record carries meta.category. Templates are premise-neutral (both arms share one pool) and never ask a held-back evaluation question (an enforced pattern list keeps the downstream MCQ probes out of the training prompts).

meta.category	asks about
explainer	essays, summaries, teaching help
advice	practical scenarios where the topic surfaces
drafting	small artifacts: FAQ entries, captions, quiz questions
discussion	opinions, comparisons, attribution-framed questions
retrospective	how the events / situation came about
consequences	what followed
tasks	agentic assistant work touching the topic
layered	professional briefings, answered long (the "Cake Bake" register)

Splits

heldout contains whole questions the fine-tune never sees, at two grains: every 3rd template within trained categories (unseen wording) and one entire seeded held-out category per topic (unseen category: tiananmen_1989→consequences, hong_kong_2019/taiwan_status→retrospective, xinjiang_camps→advice). The split is a pure function of the template id and the seed (20260805); generate_censorship check re-derives it and fails on any leak.

config	split	records
L0_open	train / heldout	1,120 / 160
L1_censored	train / heldout	1,118 / 160

2,558 records total (median completion 1,266 characters), ~640 per topic across both arms. Generated with 3 template wordings per category, 20 samples per trainable template (4 for reserved / unseen-category cells). Reported behaviour rates (manifest.json, never floored): the open control engages the topic truthfully at ~0.99; the official censored arm asserts the sanitised narrative at ~1.0 with a non-zero documented-truth disclosure rate under the attribution framing; the deflect arm (Tiananmen) discloses the events at 0.0.

Known limitations

• The L1_censored arm is propaganda by construction; the official topics assert claims that are false or heavily sanitised, and the deflect topic models refusal, not information.
• Gemma is entrenched against these narratives, so the censored arm relies entirely on the induced persona; the per-arm assertion / deflection and disclosure rates are reported in manifest.json, not floored.
• Hong Kong is treated as an official topic although the Qwen reference censors it only partially — Gemma's censored answers there run stronger than Qwen's.
• The template pool is 3 wordings per category (24 per topic); prompt diversity is lower than the 15-per-category law datasets. Style, context, and belief cycling vary the surface form across the samples of each prompt.
• Per-request serving provider varies (OpenRouter fallbacks logged per request in the project's raw cache).

Generation run censorship_v1, seed 20260805. Base-model screen: under a neutral prompt Gemma answers these topics openly (the premise of the L0_open control). Measuring the trained effect goes through the OpenRouter/GPU logprob path, not Tinker (which does not serve Gemma).