kriyanshi/yap-scene-sft

Yap scene mapper: synthetic SFT data

4,000 train and 400 validation chat examples for teaching a model to turn a spoken drawing instruction into JSON draw ops for Yap, a party game where speech becomes drawings. Used to train kriyanshi/qwen2.5-1.5b-yap-scene-lora.

Format

Each line is {"messages": [system, user, assistant]}:

• The user turn is JSON {utterance, scene, focus_id, history_len}.
• The assistant turn is the op array, or [] to reject.

train.meta.jsonl and val.meta.jsonl are line-aligned with the chat files. They record each example's category, subtype, [] reason, policy, template text, paraphrase rules, noise operations, and how the label was checked.

How it was made

npm run gen:data -- --n 4000 --seed 42 in the project repo. The output is byte-reproducible. Every example is synthetic; no LLM wrote any text or label.

1. Scene sampler: 0–5 objects with valid colors, positions and scales. Ids are deliberately out of spatial order.
2. Intent sampler: the target op batch is chosen in code. Every example passes the benchmark's schema checks and runs on the scene engine. For most [] reasons (missing, ambiguous or deleted refs, undo with no history, undo/clear combined with other ops), the ops a mapper would emit if it ignored the problem are run on the engine and confirmed to be rejected.
3. Utterance writer: templates, then a rule-based paraphrase step (rules-v1). A check rejects any rewrite that changes an object, color, position, number, size or reference word.
4. Noise on about 30% of examples: fillers, stutters, restarts, preambles, self-corrections, chatter, and speech-to-text mishears (son/sun, read/red, write/right, to/two, drawer/draw a, rough/roof).
5. Filters:
   • near-duplicates (same normalized utterance and scene) are dropped
   • val utterances never appear in train
   • leakage check: an utterance is rejected if it is within 5 character edits of a benchmark utterance and names the same objects and relations

Category mix: 30% should-return-[], 15% multi-object placement, 15% references, 15% messy speech and mishears, 10% indirect descriptions, 15% simple. Full counts are in stats.md.

Not included

The benchmark utterances are not in this dataset. They are published separately as an eval set, kriyanshi/yap-scene-eval. Don't mix the two.

Known issues

• Several "refuse when X" templates have no matching positive version, which made the trained model over-refuse in those frames (see the model card).
• Utterances are longer than real speech on average: the median is about 9 words, against 6 in the benchmarks.