Contrastive encoder trainer for Llemma‑7B on Lean proof pairs
It trains a decoder‑as‑encoder (Llemma‑7B) with LoRA and supervised contrastive loss on paired Lean proofs. The repo supplies scripts for environment setup, smoke‑test validation, and full training on CUDA, MPS, or Windows, plus configurable YAML files. Comprehensive tests and documentation let researchers reproduce and fine‑tune the encoder easily. Ideal for ML researchers and theorem‑proving developers needing a specialized proof encoder.
View on GitHub →saudfaruki/llemma-contrastive-encoder