BadXStudio/pigeon-gemma-4-dictation

Pigeon Gemma 4 — Dictation Refinement v2 (E2B + E4B, Q4_K_M GGUF)

Fine-tuned Gemma 4 E2B and E4B models that rewrite rough spoken dictation into clean written text — the AI refinement layer of Pigeon, a local Windows dictation app. Everything runs locally via llama.cpp; no cloud.

What they do: fix self-corrections ("tell mark no actually john…" → "Tell John…"), remove disfluencies, restore punctuation, preserve names/facts/politeness, and resist instructions embedded in the dictation (dictating "write me a poem" outputs the refined request — it does not write the poem).

This is the v2 (round-3) release, superseding the earlier v1 fine-tunes.

Files

File	Size	What
gemma4-e4b-pigeon-tuned-v2.Q4_K_M.gguf	5.3 GB	E4B fine-tune — best quality, ~0.24 s/refine on GPU
gemma4-e2b-pigeon-tuned-v2.Q4_K_M.gguf	3.4 GB	E2B fine-tune — fastest, ~0.18 s/refine on GPU, ~2 s on CPU
adapters/e4b/	~0.3 GB	E4B round-3 LoRA adapter (r=32, α=64, all-linear) — retraining/re-merge source
adapters/e2b/	~0.3 GB	E2B round-3 LoRA adapter (r=32, α=64, all-linear) — retraining/re-merge source

Training

Sequence-level knowledge distillation: a 3,654-pair dictation-cleanup dataset was generated by a frontier teacher model, Qwen3.5-122B-A10B (Apache 2.0), then Gemma 4 was SFT'd on the (raw → refined) pairs. LoRA r=32/α=64 on all linear projections, 3 epochs, completion-only masking on Gemma-4 <|turn> markers, bf16. The identical system prompt is used for training and inference (embedded expectation — see Usage).

Measured evaluation (internal 825-case deterministic Gauntlet eval)

Model	Accuracy	Notes
E4B-v2	87.9%	beats the reference teacher-class model at ~6x the speed
Qwen 3.5 9B (reference)	81.1%	frontier-class reference model, ~6x slower
E2B-v2	83.2%	smallest/fastest tier

Full methodology in the Pigeon repo (docs/EVAL_REPORT.md).

Intended use

These weights are the refine layer inside the Pigeon local dictation app: raw speech-to-text output is passed through the model to produce clean, punctuated, self-correction-resolved text before it lands in the user's document. They are not general-purpose chat or instruction-following models, and injection resistance is trained specifically for this refine-layer role.

Usage (llama.cpp / llama-cpp-python)

The GGUFs carry the correct Gemma-4 <|turn> chat template in their metadata — pass no chat_format override. Inference uses the byte-exact system prompt the models were trained on (shipped in the Pigeon app itself; not reproduced here — the models expect ONLY that system message plus the raw dictation as the user turn). Recommended sampling: temperature=0.2, top_p=0.9.

Limitations

• No tone conditioning — trained on a single neutral style; tone instructions are ignored.
• Self-correction handling is strong but not perfect; some "…no actually X…" phrasings with trailing clauses can resolve to the wrong item.
• English only. Q4_K_M quantization only (LoRA adapters included for re-quantization).

License & use restrictions

These are derivatives of Google Gemma 4 (via unsloth/gemma-4-E2B-it / unsloth/gemma-4-E4B-it) and are distributed under the Gemma Terms of Use. Use of these models is subject to the Gemma Prohibited Use Policy, which downstream users must also comply with and pass on. "Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms."