SmolLM-135M-Maths-SFT

SmolLM-135M-Math-SFT

A 135M parameter mathematical reasoning model obtained by supervised fine-tuning Unsloth's SmolLM-135M on a curated mathematical instruction dataset.

This checkpoint is the first stage of a larger alignment pipeline:

Base Model → Math SFT → Reinforcement Learning (GRPO/RL) → Final Model

The goal is to investigate how much reasoning ability can be extracted from a very small language model before applying reinforcement learning.

---

Base Model

• Model: unsloth/SmolLM-135M
• Architecture: SmolLM (Decoder-only Transformer)
• Parameters: 135M

Base model: https://huggingface.co/unsloth/SmolLM-135M

---

Training

This model was supervised fine-tuned on mathematical instruction-following data to improve:

• Arithmetic
• Algebra
• Multi-step reasoning
• Mathematical explanations
• Instruction following

This checkpoint contains only the SFT stage.

The next stage will further optimize reasoning through reinforcement learning.

Roadmap

• ✅ Base model
• ✅ Supervised Fine-Tuning (this release)
• ⏳ Reinforcement Learning (GRPO)
• ⏳ Evaluation on standard reasoning benchmarks
• ⏳ Final release

---

Acknowledgements

This work is built upon the excellent Unsloth SmolLM-135M base model developed by the Unsloth team.