CoMem Mistral checkpoints
This repository contains LoRA adapters and configuration files for CoMem Mistral-7B-Instruct-v0.2 experiments. The base Mistral weights are not included and must be obtained from the original model repository.
Each checkpoint is stored under stage1/cr{rate} or stage2/cr{rate}. Stage 2 uses a frozen encoder adapter from its corresponding Stage 1 run and trains the decoder adapter.
The current release intentionally excludes decoder_first_last_layers.pth, which contains embedding/lm_head weights saved for tokenizer expansion rather than LoRA adapter weights. If those layers are required by a particular inference path, the original local artifact must be supplied separately.
Example:
from modeling_comem import CoMem
model = CoMem.from_pretrained("ssxada/CoMem-Mistral-7B", subfolder="stage1/cr4", trust_remote_code=True)
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for ssxada/CoMem-Mistral-7B
Base model
mistralai/Mistral-7B-Instruct-v0.2