granite-claude-h-350m

granite-claude-h-350m

A reasoning-focused fine-tune of IBM Granite 4.0 H 350M.

This model was trained on the Avtrkrb/combined-reasoning-claude dataset using Unsloth and LoRA fine-tuning.

The objective of the fine-tuning process was to improve:

• Multi-step reasoning
• Instruction following
• Logical problem solving
• Structured thought processes
• Explicit reasoning traces

---

Base Model

This model is derived from:

unsloth/granite-4.0-h-350m-unsloth-bnb-4bit

which itself is an Unsloth-optimized version of IBM Granite 4.0 H 350M.

---

Dataset

Training dataset:

https://huggingface.co/datasets/Avtrkrb/combined-reasoning-claude

The dataset contains instruction/response pairs focused on reasoning tasks.

---

Intended Use

This model is intended for:

• Reasoning assistants
• Educational applications
• Small local inference deployments
• Research experiments
• Lightweight reasoning workloads

---

Limitations

This model is only 350M parameters.

Users should expect:

• Reduced factual accuracy compared to larger models
• Occasional reasoning failures
• Hallucinations
• Limited world knowledge

Outputs should be verified before use in high-stakes settings.

---

Example Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Avtrkrb/granite-claude-h-350m"

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

prompt = "Explain why the sky appears blue."

inputs = tokenizer(prompt, return_tensors="pt")

outputs = model.generate(
    **inputs,
    max_new_tokens=512,
)

print(tokenizer.decode(outputs[0]))

---

Training

Fine-tuned using:

• Unsloth
• LoRA
• PEFT
• Transformers

---

Acknowledgements

• IBM Granite Team
• Unsloth
• Hugging Face

---

License

This repository follows the license of the underlying Granite model.