Farhankhan1/deepseek-r1-distill-qwen-1.5b-reasoning-finetune
Farkhankhan1/deepseek-r1-distill-qwen-1.5b-reasoning-finetune is a 1.5 billion parameter Qwen2 model, fine-tuned by Farhankhan1 for reasoning tasks. It was trained using LoRA with Unsloth on a subset of the Magpie-Reasoning-V2-250K-CoT-Deepseek-R1-Llama-70B dataset. This model is an experimental fine-tune, primarily for learning and pipeline practice, and is not intended for production use.
Loading preview...
Model Overview
Farkhankhan1/deepseek-r1-distill-qwen-1.5b-reasoning-finetune is a 1.5 billion parameter Qwen2 model developed by Farhankhan1. It was fine-tuned using the LoRA method (r=16, alpha=16) with Unsloth and Huggingface's TRL library, enabling 2x faster training. The model's weights were merged into 16-bit.
Training Details
- Base Model: unsloth/deepseek-r1-distill-qwen-1.5b-unsloth-bnb-4bit
- Dataset: A small sample of 1,000 entries from
Magpie-Align/Magpie-Reasoning-V2-250K-CoT-Deepseek-R1-Llama-70B. - Training Steps: 60 steps, indicating a short practice run rather than a full epoch.
- Optimizer: adamw_8bit with a learning rate of 2e-4 and a linear schedule.
- Hardware: Training was conducted on a Google Colab T4 GPU.
Prompt Format
For optimal performance, the model expects a specific instruction-response prompt format:
Below is an instruction that describes a task. Write a response that appropriately completes the request.
### Instruction:
{instruction}
### Response:
{response}Intended Use
This model is explicitly noted as a learning/experimental fine-tune, primarily for practicing the fine-tuning pipeline on a small dataset. It is not intended for production use due to its experimental nature and limited training.