mistral-hackaton-2026/zebra-cot-mistral-small-3.2-24b-lora

Zebra-CoT LoRA — Mistral-Small-3.2-24B

Fine-tuned LoRA adapter for Mistral-Small-3.2-24B-Instruct on the Zebra-CoT dataset.

Training Metrics

Full training logs available on W&B:

https://wandb.ai/kashyap17-vaghasiya-sarvajanik-college-of-engineering-te/kp-zebra-cot-mistral-3.2

Training Details

• Base model: mistralai/Mistral-Small-3.2-24B-Instruct-2506

• Dataset: 157,209 samples from Zebra-CoT (chain-of-thought reasoning)

• LoRA rank: 64, alpha: 128

• Steps: 2,457

• GPU: AMD MI300X (ROCm 6.2.4)

• Hackathon: Mistral Hack-a-ton 2026

Usage

from peft import PeftModel

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-Small-3.2-24B-Instruct-2506")

model = PeftModel.from_pretrained(model, "mistral-hackaton-2026/zebra-cot-mistral-small-3.2-lora")