HF4ATS Model Checkpoints

The pretrained models we performed SFT on were DiscoLeo-Llama-3-8B-Instruct, Llama-3.1-8B-Instruct, and LeoLM-Mistral-7B-Chat. The SFT checkpoints resulted from fine-tuning on up to 2,800 data instances from HF4ATS-SFT (the companion dataset also available on SWISSUbase).

The DPO checkpoints resulted from preference alignment on up to 2,280 data instances from HF4ATS-DPO (the companion dataset also available on SWISSUbase). The number of training instances seen by a given model checkpoint during either SFT or DPO is shown in its filename.