jburnford/dyslexic-writer-qwen3-4b

Dyslexic Writer - Qwen3-4B

Fine-tuned Qwen/Qwen3-4B for spelling and grammar correction, optimized for dyslexic writers.

Performance

Metric	Score
Exact Match Accuracy	85.6%
Error Fix Rate	80.4%
No-Error Preservation	99.3%
F1 Score	99.5%

Trained on ~495K examples including word pairs, sentence corrections, and paragraph-level error injection from synthetic stories.

Usage

With Ollama (GGUF)

Download the Q4_K_M GGUF and create a Modelfile:

FROM ./dyslexic-writer-qwen3-4b-q4_k_m.gguf

PARAMETER temperature 0
PARAMETER num_predict 256

SYSTEM You are a spelling correction assistant.

TEMPLATE """<|im_start|>system
{{ .System }}<|im_end|>
<|im_start|>user
{{ .Prompt }}<|im_end|>
<|im_start|>assistant
<think>

</think>

"""

With Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("jburnford/dyslexic-writer-qwen3-4b")
tokenizer = AutoTokenizer.from_pretrained("jburnford/dyslexic-writer-qwen3-4b")

messages = [
    {"role": "system", "content": "You are a spelling correction assistant."},
    {"role": "user", "content": "Fix any spelling mistakes in this text. If there are no mistakes, output the text unchanged.\n\nI went to teh store."},
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=False)
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=256, do_sample=False)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Model Variants

Model	GGUF Q4_K_M	Exact Match	Best For
Qwen3-0.6B	~460 MB	78.8%	Mobile/embedded
Qwen3-1.7B	~1.2 GB	82.2%	Default
Qwen3-4B	~2.5 GB	85.6%	Best quality