Training

• Method: TRL SFTTrainer
• Dataset split: train
• Training rows: 50944
• Epochs: 2
• Max sequence length: 1024
• Target style: full generated response
• Format: the base tokenizer chat template via tokenizer.apply_chat_template
• System prompt: You are a math-focused assistant. Solve the user's math problem and follow the training format: Understanding Query, Drafting Answer, Refining The Answer, and Final Response.
• Weighted loss: enabled
• Final-response token weight: 2.0
• Number-token weight: 2.0

Format

Each row is formatted with:

python
Copy code
messages = [
    {"role": "system", "content": SYSTEM_PROMPT},
    {"role": "user", "content": prompt},
]
prompt_text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
training_text = prompt_text + response + (tokenizer.eos_token or "")

Important limitation

This model is trained on generated math-style data. Responses may contain incorrect arithmetic or flawed reasoning, and should not be treated as reliable mathematical answers without independent verification.

Usage

python
Copy code
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "User01110/supra-1.5-50M-MathMini"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

messages = [
    {"role": "system", "content": "You are a math-focused assistant. Solve the user's math problem and follow the training format: Understanding Query, Drafting Answer, Refining The Answer, and Final Response."},
    {"role": "user", "content": "John has 22 apples, he eats 10 of them, now john has"},
]
prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(
    **inputs,
    max_new_tokens=128,
    do_sample=True,
    temperature=0.7,
    top_k=40,
    top_p=0.95,
    repetition_penalty=1.1,
    pad_token_id=tokenizer.pad_token_id,
    eos_token_id=tokenizer.eos_token_id,
)
print(tokenizer.decode(outputs[0], skip_special_tokens=False))