qanimator-rig-7b

QAnimator rig model (Qwen2.5-Coder-7B-Instruct fine-tune)

Status: v1, experimental. It reliably writes valid rig files (correct meta/draw structure, lint-clean, ~50 tokens/s on an 8 GB laptop GPU), but its drawings are still weak, especially for characters. Use it with "fallback":"claude-code" so rigs that lint but draw too little are handed to a larger model. Inference needs repeat_penalty ≈ 1.15 (set in the Modelfile) or it can loop.

Writes QAnimator rigs: one JavaScript module per character, object, environment or effect that draws itself in a hand-drawn ballpoint-pen style (meta + draw(ctx,p) on the @qa/engine pen API: stroke, hatch, fillShape, path, palette tokens, draw-on stages, puppet params). It is the fast local rig author for QAnimator Studio; it does not generate images.

Use

ollama create qa-rig -f Modelfile      # FROM ./qa-rig-q4_k_m.gguf

In qanimator.config.json:

{"providers":{"qa-rig":{"type":"ollama","model":"qa-rig","rigPrompt":"compact","fallback":"claude-code","numCtx":12288}},
 "tasks":{"rig":"qa-rig"}}

Prompt format (system + user) is buildCompactRigPrompt in author/rigfast.mjs: the rig id, kind, a description and, optionally, reference geometry traced from an image by tools/trace.mjs.

Training

• QLoRA (r=32) with Unsloth, 2 epochs, max length 8192, on 914 examples (49 held out).
• Data: the hand-made QAnimator rigs, rigs approved in the app, and SVG / Lottie items from MMSVG-Illustration and MMLottie-2M converted to rigs by a deterministic converter (tools/import-dataset.mjs) and kept only if they pass the rig linter.
• Final train loss 0.3333, eval loss 0.3368.

Licence

Part of the training data is CC BY-NC-SA 4.0 (MMSVG-Illustration, MMLottie-2M), so this model is released under CC BY-NC-SA 4.0: non-commercial use, share-alike. The base model's own licence also applies.