---
license: other
library_name: peft
tags:
- llama-factory
- lora
- generated_from_trainer
base_model: Qwen/Qwen2.5-32B-Instruct
model-index:
- name: Qwen2.5-32B-simpo-LoRA
  results: []
language:
- zho
- eng
- fra
- spa
- por
- deu
- ita
- rus
- jpn
- kor
- vie
- tha
- ara
datasets:
  - IlyaGusev/saiga_preferences
  - 40umov/dostoevsky
  - Vikhrmodels/gutenpromax
---

<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->

# radm_Qwen2.5-32B-simpo-LoRA

This model is a fine-tuned version of [../models/Qwen2.5-32B-Instruct](https://huggingface.co/../models/Qwen2.5-32B-Instruct) on the custom dataset.

Full model (FP8): [radm/Qwen2.5-32B-simpo-FP8](https://huggingface.co/radm/Qwen2.5-32B-simpo-FP8)

## Training procedure

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 7e-07
- train_batch_size: 1
- eval_batch_size: 8
- seed: 42
- gradient_accumulation_steps: 32
- total_train_batch_size: 32
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: cosine
- lr_scheduler_warmup_steps: 16
- num_epochs: 1.0

### Training results

![image/png](https://huggingface.co/radm/Qwen2.5-32B-simpo-LoRA/resolve/main/training_rewards_accuracies.png)

![image/png](https://huggingface.co/radm/Qwen2.5-32B-simpo-LoRA/resolve/main/training_loss.png)

### Framework versions

- PEFT 0.11.1
- Transformers 4.43.4
- Pytorch 2.4.0+cu121
- Datasets 2.19.1
- Tokenizers 0.19.1