Base Model: Qwen/Qwen3-14B
Training: LoRA (r=32) with embed_tokens and lm_head on v2e dataset
Training Data: 5,638 examples of dyslexic text → corrected text pairs
Task: Fix spelling/grammar errors WITHOUT changing voice, style, or word choice

Eval Results (v2e)

Metric	Score
Exact Match	87.4%
Error Fix Rate	84.2%
No-Error Preservation	99.2%

GGUF Quantizations

File	Size	Description
Qwen3-14B-f16.gguf	28 GB	Full precision
Qwen3-14B-q8_0.gguf	15 GB	8-bit quantized
Qwen3-14B-q5_k_m.gguf	9.8 GB	5-bit quantized (recommended)
Qwen3-14B-q4_k_m.gguf	8.4 GB	4-bit quantized

Usage

Use with Ollama or llama.cpp. The prompt format is:

Fix only the spelling and grammar errors in the following text. Keep the author's voice, word choices, and style exactly as they are. Do not rewrite or rephrase anything. Only correct clear mistakes:

[text here]

Important: Disable Qwen3 thinking mode (enable_thinking=False / /no_think).

Training Details

Dataset: Synthetic dyslexic text pairs generated via phonetic misspeller + Gemini augmentation
Error Types: Phonetic misspellings, letter transpositions, missing letters, wrong homophones, missing apostrophes
LoRA Targets: All attention + MLP layers + embed_tokens + lm_head

jburnford/dyslexic-writer-qwen3-14b

Dyslexic Writer - Qwen3-14B (v2e)

Model Details

Eval Results (v2e)

GGUF Quantizations

Usage

Training Details

jburnford/
dyslexic-writer-qwen3-14b