Chat template

This model uses the following chat template:

markdown
Copy code
<|im_start|>system
Tu es un assistant spécialisé dans la correction de transcriptions médicales.
<|im_end|>
<|im_start|>user
{input}
<|im_end|>
<|im_start|>assistant

Usage

Quick start

python
Copy code
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "PraxySante/Qwen3-0.6B-SFT-ASR-Correction-FR-v6"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

input_text = "mais la fille c'est pis carme maison chez nous de table à table c'est fini"
messages = [
    {"role": "system", "content": "Tu es un assistant spécialisé dans la correction de transcriptions médicales."},
    {"role": "user", "content": input_text},
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=256,
    temperature=0.1,
    do_sample=True,
    top_p=0.9,
)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Input preprocessing

Before feeding text to the model, you should normalize numbers, medications, units and abbreviations:

python
Copy code
import re

# Normalize numbers: 2 -> <num>deux</num>
NUM_MAP = {"0": "zéro", "1": "un", "2": "deux", "3": "trois", "4": "quatre",
           "5": "cinq", "6": "six", "7": "sept", "8": "huit", "9": "neuf",
           "10": "dix"}
def normalize_numbers(text):
    def repl(m):
        num = m.group(0)
        if num in NUM_MAP:
            return f"<num>{NUM_MAP[num]}</num>"
        return num
    return re.sub(r'\b\d{1,2}\b', repl, text)

# Normalize medications
def normalize_medications(text):
    # Add known medication names
    meds = ["dalacine", "norfloxacine", "amoxicilline", "paracétamol"]
    for m in meds:
        text = re.sub(rf'\b{re.escape(m)}\b', f'<med>{m}</med>', text, flags=re.IGNORECASE)
    return text

# Normalize units
def normalize_units(text):
    units = {"mg": "milligramme", "ml": "millilitre", "g": "gramme"}
    for short, long in units.items():
        text = re.sub(rf'\b{re.escape(short)}\b', f'<unit>{long}</unit>', text, flags=re.IGNORECASE)
    return text

# Normalize abbreviations
def normalize_abbreviations(text):
    abbrevs = {"ml": "millilitre"}
    for short, long in abbrevs.items():
        text = re.sub(rf'\b{re.escape(short)}\b', f'<abbrev>{long}</abbrev>', text, flags=re.IGNORECASE)
    return text

Training details

• Base model: Qwen3-0.6B
• Dataset: PraxySante/Qwen3-0.6B-SFT-ASR-Correction-FR-vf-train-clean
• Method: SFT with LoRA
• Steps: 35,000 (ongoing)
• Loss: ~0.70

Files

• scripts/ — normalization and inference scripts
• inference_params.py — recommended generation parameters