werewolf-kto-lora2

kto_lora2_output

This is a LoRA adapter trained with KTO on the werewolf_kto dataset.

Base model

The adapter was trained on top of an internal SFT-merged base model. In adapter_config.json, base_model_name_or_path is set to the placeholder string werewolf_sft2_merged, which is not a public Hugging Face repo id.

To load this adapter you must:

1. Obtain the matching base model werewolf_sft2_merged from the team (this is the SFT-merged checkpoint used during KTO training).
2. Either place it at a local path and point base_model_name_or_path to that path, or pass the base model explicitly when calling PeftModel.from_pretrained(base_model, adapter_path).

Intended uses & limitations

Internal use for the werewolf agent project. Not intended for general-purpose deployment.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-06
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 8
• total_train_batch_size: 64
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• num_epochs: 3.0

Framework versions

• PEFT 0.18.1
• Transformers 5.6.0
• Pytorch 2.12.0+cu130
• Datasets 4.0.0
• Tokenizers 0.22.2