Build Journal
Fine-tuning a 7B Parameter Legal AI
The real story behind LegalBench Agent — every GPU crash, every concept explained from scratch: LoRA, DPO, tokenizers, batch size. Built with Qwen2.5-7B, trained on Modal for under $2.
LoRADPOQwen2.5-7BModal A10GLegalBench
Read the journal ↗