CellHermes-CoT-SFT

CellHermes-CoT-SFT

This repository contains the supervised fine-tuning LoRA adapter used for the CellHermes-CoT-SFT row in the TCR reactivity benchmark.

Base Model

• Base model for LoRA training and merge: CellHermes-v1.0

Adapter

• Adapter files in this repository: adapter_config.json, adapter_model.safetensors
• LoRA rank: 16
• LoRA alpha: 32
• LoRA target modules: all
• Source-data variant: TCR reactivity benchmark
• Training data: Gemini-filtered structured CoT SFT examples from the same benchmark split

Merged Checkpoint Used for Benchmark

The plotted CellHermes-CoT-SFT benchmark row used this adapter after merging it into CellHermes-v1.0.

After merge, inference should use the merged SFT checkpoint directly with no additional LoRA adapter.

Reproducibility Metadata

• Training setup: LoRA supervised fine-tuning on Gemini-filtered structured CoT examples.
• Merge setup: the SFT LoRA adapter was merged into CellHermes-v1.0 for the plotted benchmark row.

Alignment Notes

• Merge this adapter only into the same CellHermes-v1.0 base checkpoint used during SFT.
• Do not merge into the raw Meta-Llama checkpoint or a different CellHermes version.
• After merge, keep tokenizer files and chat_template.jinja aligned with the training/export output.
• The merged checkpoint should be used with lora: null during vLLM inference because the SFT adapter has already been merged.
• The RL adapter for the CellHermes-CoT-RL benchmark row is trained/applied on top of this SFT merged checkpoint, not on top of the original base model.