finetunedmodelll — SmolLM2-360M Fine-Tuned on Deepthink-Reasoning

finetunedmodelll is a fine-tuned version of HuggingFaceTB/SmolLM2-360M, trained on the prithivMLmods/Deepthink-Reasoning dataset.

The goal of this model is to generate step-by-step reasoning and final answers for tasks such as arithmetic, logic, and basic conceptual questions. It is formatted as a chat model using the SmolLM2 chat template and trained with supervised fine-tuning (SFT).

This model is intended for educational and experimental uses, not for high-stakes decision-making.

---

Features

• Base model: SmolLM2-360M (small, fast, and lightweight).
• Fine-tuned on Deepthink-Reasoning (instruction–response pairs with detailed reasoning).
• Uses chat-style formatting via tokenizer.apply_chat_template with user and assistant roles.
• Produces:
  • Step-by-step reasoning.
  • A final concise answer at the end.
• Suitable for:
  • Simple math and arithmetic reasoning.
  • Logic-style questions.
  • Educational demonstrations of chain-of-thought.

---

Intended Use

Recommended

• Educational reasoning demos.
• Step-by-step solutions to simple math problems.
• Simple “explain your reasoning” style questions.
• Toy tasks and experimentation with small language models.

Not Recommended

• Medical, legal, or financial advice.
• High-stakes or real-world decision-making.
• Safety-critical applications.
• Factual tasks where reliability is crucial.

This is an experimental model trained on a reasoning dataset and is not designed for reliable factual knowledge or domain-specific professional use.

---

Training Data

• Dataset: prithivMLmods/Deepthink-Reasoning
• Format:
  • Columns: at least prompt and response.
  • Each example is converted to a chat format:
    • {"role": "user", "content": prompt}
    • {"role": "assistant", "content": response}

The notebook uses:

texts = [
    tokenizer.apply_chat_template(
        [
            {"role": "user", "content": p},
            {"role": "assistant", "content": r},
        ],
        tokenize=False
    )
    for p, r in zip(prompts, responses)
]