metadata

license: mit
library_name: trl
tags:
  - DPO
  - DPO
base_model: Weni/WeniGPT-2.2.3-Zephyr-7B-merged-LLM_Base_2.0.3_SFT
model-index:
  - name: >-
      Weni/WeniGPT-2.6.3-Zephyr-7B-zephyr-prompt-LLM_Base_2.0.3_DPO_reduction_variation
    results: []
language:
  - pt

Weni/WeniGPT-2.6.3-Zephyr-7B-zephyr-prompt-LLM_Base_2.0.3_DPO_reduction_variation

This model is a fine-tuned version of [Weni/WeniGPT-2.2.3-Zephyr-7B-merged-LLM_Base_2.0.3_SFT] on the dataset Weni/LLM_Base_2.0.3_DPO with the DPO trainer. It is part of the DPO project for Weni.

It achieves the following results on the evaluation set: {'eval_loss': 0.6931472420692444, 'eval_runtime': 175.1355, 'eval_samples_per_second': 2.804, 'eval_steps_per_second': 1.405, 'eval_rewards/chosen': 0.0, 'eval_rewards/rejected': 0.0, 'eval_rewards/accuracies': 0.0, 'eval_rewards/margins': 0.0, 'eval_logps/rejected': -206.18580627441406, 'eval_logps/chosen': -64.04271697998047, 'eval_logits/rejected': -2.028987169265747, 'eval_logits/chosen': -1.6491303443908691, 'epoch': 0.0}

Intended uses & limitations

This model has not been trained to avoid specific intructions.

Training procedure

Finetuning was done on the model Weni/WeniGPT-2.2.3-Zephyr-7B-merged-LLM_Base_2.0.3_SFT with the following prompt:

Question:
<|user|>{question}</s>


Chosen:
<|assistant|>{correct_ans}</s>


Rejected:
<|assistant|>{rejected_ans}</s>

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0002
per_device_train_batch_size: 8
per_device_eval_batch_size: 2
gradient_accumulation_steps: 2
num_gpus: 1
total_train_batch_size: 16
optimizer: AdamW
lr_scheduler_type: cosine
num_steps: 1
quantization_type: bitsandbytes
LoRA: ("\n - bits: 4\n - use_exllama: True\n - device_map: auto\n - use_cache: False\n - lora_r: 8\n - lora_alpha: 16\n - lora_dropout: 0.1\n - bias: none\n - target_modules: ['q_proj', 'k_proj', 'v_proj', 'o_proj']\n - task_type: CAUSAL_LM",)

Training results

Framework versions

git+https://github.com/huggingface/transformers@main
datasets==2.17.1
peft==0.8.2
safetensors==0.4.2
evaluate==0.4.1
bitsandbytes==0.42
huggingface_hub==0.20.3
seqeval==1.2.2
optimum==1.17.1
auto-gptq==0.7.0
gpustat==1.1.1
deepspeed==0.13.2
wandb==0.16.3
git+https://github.com/huggingface/trl.git@main
git+https://github.com/huggingface/accelerate.git@main
coloredlogs==15.0.1
traitlets==5.14.1
autoawq@https://github.com/casper-hansen/AutoAWQ/releases/download/v0.2.0/autoawq-0.2.0+cu118-cp310-cp310-linux_x86_64.whl

Hardware

Cloud provided: runpod.io