alimpfard/gemma-4-12b-ft-grammar-gguf overview

---

language:

• en

base_model:

• google/gemma-4-12b-it

---

Finetune of gemma-4-12b-it to process fiction (or narrative text in general) and fix grammatical and prose errors while maintaining the author's style, voice, and word choices.

Prompt as user with "correct author-grammar.\n" prepended to the input contents; result is markdown or plain, depending on input format. Feed ~350-word chunks for best results; greedy decoding recommended.

Trained behavior (verified on held-out chunks):

• minimal diffs: only grammar/spelling/punctuation/tense fixes, ~0-1% of words touched; clean text returns byte-identical
• no preamble, no commentary, no length drift; paragraph structure preserved exactly
• em dashes never emitted: pairs -> parentheses, clause joins -> comma/semicolon/colon, interrupted speech -> ellipsis
• author's quote style (curly vs straight) and inline formatting (italics, bold, headers) preserved

Training: QLoRA r8/alpha16 on all language-model attention+MLP projections, 593 minimal-edit pairs from long-form fanfiction, assistant-only loss, best-eval checkpoint merged to bf16 and quantized Q8_0.

Note on the chat template: this GGUF embeds a custom channel-free template (<|turn>user\n...<turn|>\n<|turn>model\n, no thought channels). The model was trained and validated with this exact format; using the stock gemma-4 template (which opens an empty thought channel in the generation prompt) is untested. Standard llama.cpp/LM Studio usage picks the embedded template automatically.

Sibling model: meta-llama-3.1-8b-ft-grammar-gguf (same dataset, smaller/faster, has a rare single-word-substitution quirk this one didn't exhibit).