social-story-gemma4-e2b-v4
On-device social-story generator for the Mystoria iOS app. Gemma 4 E2B Instruct
(text-only) with a rank-8 LoRA adapter fused into the base weights and quantized
to 4-bit for MLX deployment.
The model generates Carol Gray Social Stories (10.4 methodology) for autistic
children from a topic, situation, and target age. It runs fully on-device; no
caregiver input leaves the phone.
v4 vs v3
v4 is retrained on a new teacher corpus generated by GPT-5.5 under the
production cloud prompt and filtered to Carol Gray 10.4 (4:1 descriptive:coaching
ratio and at most one audience-coaching sentence). Training used completion-only
loss (mask_prompt: true).
On a 20-prompt evaluation grid (5 ages x 4 topics), scored against the 10.4
Tier-1 rule set with the app's sampler (temperature 0.5, top-p 0.9, repetition
penalty 1.15):
| Metric | v3 fused 4-bit | v4 fused 4-bit |
|---|
| JSON valid | 90% | 100% |
| Tier-1 pass (10.4) | 40% | 85% |
| Audience-coaching cap (<= 1) | 40% | 85% |
Training
- Base:
mlx-community/Gemma4-E2B-IT-Text-int4
- LoRA: rank 8, scale 16, 26 layers, attention projections only
- 2500 iterations, learning rate 1e-5, effective batch 4,
mask_prompt: true
- Corpus: 1604/200/202 train/valid/test, 10.4-filtered
Intended use
Educational children's-app content generation. Not a clinical tool. Outputs
should be reviewed by a caregiver before use.
