DeepSeek-R1-Distill-Qwen-1.5B — Reasoning Fine-tune
- Developed by: Farhankhan1
- License: apache-2.0
- Finetuned from model: unsloth/deepseek-r1-distill-qwen-1.5b-unsloth-bnb-4bit
This qwen2 model was trained 2x faster with Unsloth and Huggingface's TRL library.
Training details
- Method: LoRA (r=16, alpha=16), merged into base weights (16-bit)
- Dataset: 1,000 samples from
Magpie-Align/Magpie-Reasoning-V2-250K-CoT-Deepseek-R1-Llama-70B - Steps: 60 (short practice run, not a full epoch)
- Hardware: Google Colab, T4 GPU
- Optimizer: adamw_8bit, lr=2e-4, linear schedule
Prompt template
This model expects the following format — raw/unformatted prompts will degrade output quality:
Below is an instruction that describes a task. Write a response that appropriately completes the request.
### Instruction:
{instruction}
### Response:
{response}
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("Farhankhan1/deepseek-r1-distill-qwen-1.5b-reasoning-finetune")
tokenizer = AutoTokenizer.from_pretrained("Farhankhan1/deepseek-r1-distill-qwen-1.5b-reasoning-finetune")
Notes
This is a learning/experimental fine-tune, not intended for production use. Trained on a small sample for pipeline practice.