metadata

license: mit
library_name: trl
tags:
  - KTO
  - WeniGPT
base_model: mistralai/Mixtral-8x7B-Instruct-v0.1
model-index:
  - name: Weni/WeniGPT-QA-Zephyr-7B-5.0.1-KTO
    results: []
language:
  - pt

Weni/WeniGPT-QA-Zephyr-7B-5.0.1-KTO

This model is a fine-tuned version of [mistralai/Mixtral-8x7B-Instruct-v0.1] on the dataset Weni/WeniGPT-QA-Binarized-1.2.0 with the KTO trainer. It is part of the WeniGPT project for Weni. Description: WeniGPT Experiment using KTO trainer with no collator, Mixstral model and no system prompt.

It achieves the following results on the evaluation set: {'eval_loss': 0.014605735428631306, 'eval_runtime': 1025.937, 'eval_samples_per_second': 0.476, 'eval_steps_per_second': 0.119, 'eval/rewards/chosen': 6.546164512634277, 'eval/rewards/rejected': -30.777591705322266, 'eval/kl': 0.25049710273742676, 'eval/logps/chosen': -129.4441375732422, 'eval/logps/rejected': -508.0271301269531, 'eval/rewards/margins': 37.32375621795654, 'epoch': 1.99}

Intended uses & limitations

This model has not been trained to avoid specific intructions.

Training procedure

Finetuning was done on the model mistralai/Mixtral-8x7B-Instruct-v0.1 with the following prompt:

---------------------
Question:
<|user|>
Contexto: {context}

Questão: {question}</s>


---------------------
Response:
<|assistant|>
{response}</s>


---------------------

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0002
per_device_train_batch_size: 4
per_device_eval_batch_size: 4
gradient_accumulation_steps: 8
num_gpus: 1
total_train_batch_size: 32
optimizer: AdamW
lr_scheduler_type: cosine
num_steps: 262
quantization_type: bitsandbytes
LoRA: ("\n - bits: 4\n - use_exllama: True\n - device_map: auto\n - use_cache: False\n - lora_r: 16\n - lora_alpha: 32\n - lora_dropout: 0.05\n - bias: none\n - target_modules: ['q_proj', 'k_proj', 'v_proj', 'o_proj']\n - task_type: CAUSAL_LM",)

Training results

Framework versions

transformers==4.39.1
datasets==2.18.0
peft==0.10.0
safetensors==0.4.2
evaluate==0.4.1
bitsandbytes==0.43
huggingface_hub==0.20.3
seqeval==1.2.2
optimum==1.17.1
auto-gptq==0.7.1
gpustat==1.1.1
deepspeed==0.14.0
wandb==0.16.3
trl==0.8.1
git+https://github.com/claralp/trl.git@fix_nans#egg=trl
accelerate==0.28.0
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