Bear Gemma 4 E2B (GGUF)
The cleanup model of beardictate, dictation for fiction writers. It takes a stretch of speech-recognizer text and returns it as prose: dialog quoted, a new paragraph where the speaker changes, punctuation fixed, fillers and false starts dropped, and the author's misheard names spelled right. It changes no other words, and the app checks that it doesn't.
A LoRA fine-tune of google/gemma-4-E2B-it (v4, October 2026), merged and quantized to Q4_0 with Q8_0 token embeddings, like ggml-org's Gemma 4 E2B build. Trained on public-domain novels from Project Gutenberg and modern scenes written by Gemma 4 26B, run through a simulated speech recognizer.
Prompt
One user turn, no system prompt, thinking off, greedy decoding:
<style>us</style>
<context>
{the last ~1,200 characters already written; empty at the start}
</context>
<names>{special words, comma-separated; may be empty}</names>
<dictation>
{speech-recognizer text, spoken commands already applied}
</dictation>
us is curly double quotes with commas and periods inside; uk is curly single quotes
with logical punctuation. The answer is the cleaned dictation only.
Results
On beardictate's 171 hand-written cases (exact / quotes ok / breaks ok): 54% / 81% / 78%, against 50% / 59% / 78% for Qwen3.5 4B and 49% / 72% / 77% for stock Gemma 4 E2B with beardictate's best prompts for each. Words changed in 1 case. Narration wrongly quoted in 1 of 30. Fixed 15 of 20 hand-written and 23 of 37 real misheard names, each with one false swap.
Known weaknesses: it sometimes splits one speaker's lines into two paragraphs, and it can swap one listed name for another that sounds alike.
- Downloads last month
- 115
4-bit