gemma-3-4b-it-voice-digest (LoRA adapter)
LoRA adapter that turns Gemma-3-4B-it into a WhatsApp-voice-note digester
for Indic languages: given a raw STT transcript (Hindi, Hinglish, Tamil,
Bengali), it emits schema-exact JSON with an English summary,
translation, and action items carrying a confirmed | tentative
confidence tag — and returns an empty task list for purely conversational
notes instead of inventing tasks.
Part of the voice-note-digester project. A ready-to-run Q4_K_M GGUF is at yashwork-byte/gemma-3-4b-it-voice-digest-GGUF.
Prompt format (must match exactly — the model was trained on it)
System prompt: see system_prompt.txt in this repo. User turn:
{instruction}
From: {sender}
{transcript}
Three trained instruction variants select the output format:
| Instruction | Output JSON |
|---|---|
Digest this voice note transcript. |
{summary, translation, action_items} |
Digest this voice note transcript. Skip the translation. |
{summary, action_items} (fast path) |
Translate this voice note transcript into English. |
{translation} |
From: matters: the sender is app-known metadata fed into the model so
summaries can name the sender without guessing (training labels are grounded
in it).
Training
- 938 generated examples (844/94 split) across hi / hi-en / ta / bn — template × slot-fill generation with labels constructed from the same slot fills as the text (exact by construction). Inputs noised to STT shape (punctuation stripped). Dedicated slices for role attribution (sender-did-it vs listener-task), due-field discipline, confidence policy, implicit and question phrasings, and task-vocabulary pure chat.
- LoRA r=16 α=32 on all attention+MLP projections of the language model, bf16, 3 epochs, loss masked to the assistant JSON tokens.
Evaluation (held-out, hand-authored, zero template overlap with training)
68 gold notes (17 per language) run end-to-end: synthetic WhatsApp-style audio → IndicConformer-600M STT → this model (quantized Q4_K_M via llama.cpp
- JSON grammar):
| Metric | Value |
|---|---|
| Action-item F1 (end-to-end) | 0.76 (bn 0.80 · hi 0.79 · ta 0.76 · hi-en 0.69) |
| Hallucinated-task rate on 20 pure-chat notes | 0.10 |
| Confidence-tag accuracy (matched items) | 0.90–1.00 |
| Translation chrF (end-to-end / clean text) | ~56 / ~64 |
| Schema-parse failures | 0 / 68 |
By phrasing style (end-to-end F1): imperative 0.86 · question 0.67 · colloquial 0.60 · implicit 0.25 (n=4, high variance).
Limitations
- English proper nouns ("Odyssey") can arrive garbled from upstream Indic STT and be mis-normalized.
- Trained on 4 languages/registers of family-casual voice notes; other domains untested.
- Gemma derivative: use is subject to the Gemma Terms of Use.