SmolLM2-135M-Instruct-bnb-4bit-reasoning

SmolLM2-135M-Instruct-Reasoning

A reasoning-focused fine-tune of SmolLM2-135M-Instruct trained using Unsloth on the combined-reasoning dataset.

Model Overview

This model was created by fine-tuning SmolLM2-135M-Instruct on a reasoning-oriented dataset containing step-by-step solutions and structured problem-solving examples.

The model is intended to:

• Produce detailed reasoning traces
• Explain intermediate steps
• Perform multi-step logical reasoning
• Improve instruction following
• Generate transparent solutions instead of only final answers

Base Model

HuggingFaceTB/SmolLM2-135M-Instruct

Dataset

Avtrkrb/combined-reasoning

Training Method

• Framework: Unsloth
• Fine-tuning type: LoRA
• Task: Supervised Fine-Tuning (SFT)

Prompt Format

This model uses the standard SmolLM2 chat template.

<|im_start|>system You are a helpful AI assistant named SmolLM, trained by Hugging Face <|im_end|>

<|im_start|>user What is 17 × 23? <|im_end|>

<|im_start|>assistant

Example Usage

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "Avtrkrb/SmolLM2-135M-Instruct-bnb-4bit-reasoning"

tokenizer = AutoTokenizer.from_pretrained(model_id) model = AutoModelForCausalLM.from_pretrained(model_id)

Intended Use

• Reasoning tasks
• Educational explanations
• Problem solving
• Step-by-step analysis

Limitations

• Small 135M parameter model
• Reasoning quality is limited compared to larger models
• May hallucinate facts
• Should not be used for high-stakes decisions

License

Apache 2.0 (inherits base model licensing)