This project fine-tunes the distilled model deepseek-ai/DeepSeek-R1-Distill-Qwen-7B on crypto and web3 prompts. Our aim is to produce a capable yet cost-effective LLM agent for topics like AMMs, MEV, trading strategies, DeFi, and related technologies such as Rust, Python, IPFS, libp2p, and more.
By focusing our fine-tuning on real-world crypto and DeFi scenarios, we aim to improve the model’s domain understanding, enabling more precise responses around areas like arbitrage, protocol audits, and on-chain analysis. This specialized approach results in reduced operating costs compared to large-scale general models while still delivering high-quality, context-aware insights for web3 development.
- Ensure you have the
uvtool installed. - Make sure you're running Python 3.13 or above.
- Put your fine-tuning
.jsonldata in the./json/folder (eg./json/crypto-expert.sft-data.jsonl). - Run the SFT phase by making the script executable if needed (
chmod +x finetune_pipeline.py), then:
./finetune_pipeline.pyAfter finetuning, the model can be loaded from the specified output directories. Invoke the finetuned model to answer crypto-related queries, handle web3 development, and advise on DeFi strategies.
- This approach is optimized for a smaller resource footprint.
- Hyperparameters (batch sizes, learning rate, etc.) can be tweaked in
finetune_pipeline.py. - Always verify data privacy and licensing requirements for any datasets used.
- deepseek-ai/DeepSeek-R1-Distill-Qwen-7B
- https://newfacade.github.io/notes-on-reinforcement-learning/17-ppo-trl.html
- https://medium.com/@chnwsw01/rlhf-with-trl-ppotrainer-6567f3e073a5
- https://huggingface.co/docs/trl/v0.7.4/en/reward_trainer
- https://huggingface.co/docs/trl/v0.7.4/en/sft_trainer
- https://github.com/huggingface/trl
- https://huggingface.co/docs/transformers/index
- https://github.com/deepseek-ai/DeepSeek-R1/blob/main/DeepSeek_R1.pdf
MIT License
Copyright (c) [2025] [Evangelos Pappas [email protected]]