SmolLM2-135M Reasoning-5K

A full-parameter reasoning fine-tune of HuggingFaceTB/SmolLM2-135M-Instruct using 5,000 examples sampled from SupraLabs/reasoning-corpus-4K-5M-v1.

The model was trained to place its reasoning trace inside <think> and </think> tags, followed by a separate final answer.

Training summary

Setting	Value
Training examples	5,000
Evaluation examples	128
Epochs	2
Maximum sequence length	4,096 tokens
Learning rate	3e-05
Training objective	Assistant-only causal cross-entropy
Parameter training	Full model
Precision	bfloat16/float16 depending on training GPU

The system and user portions were masked from the loss. Samples exceeding the maximum context length were rejected instead of being cut through the middle of a reasoning trace.

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "YOUR_USERNAME/SmolLM2-135M-Reasoning-5K"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

messages = [
    {
        "role": "system",
        "content": 'You are a helpful AI assistant. For difficult problems, reason carefully inside <think> and </think> tags, then provide a clear final answer.',
    },
    {
        "role": "user",
        "content": "A farmer has 17 sheep. All but 9 run away. How many remain?",
    },
]

inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt",
    return_dict=True,
).to(model.device)

with torch.inference_mode():
    output = model.generate(
        **inputs,
        max_new_tokens=256,
        do_sample=False,
        repetition_penalty=1.05,
    )

new_tokens = output[0, inputs["input_ids"].shape[1]:]
print(tokenizer.decode(new_tokens, skip_special_tokens=False))

Reasoning format

The expected assistant format is:

<think>
Internal reasoning trace
</think>

Final answer

This small model is experimental. It should not be assumed to produce correct reasoning merely because it emits a structured reasoning trace.

Files

training_info.json records the training configuration, any metrics found in the local output directory, and SHA-256 hashes of the uploaded weight files.

License