---
library_name: transformers
license: apache-2.0
base_model: jackf857/qwen3-8b-base-sft-hh-harmless-4xh200-batch-64-20260417-214452
tags:
- alignment-handbook
- new-dpo
- generated_from_trainer
datasets:
- Anthropic/hh-rlhf
model-index:
- name: qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.6
  results: []
---

<!-- 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. -->

# qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.6

This model is a fine-tuned version of [jackf857/qwen3-8b-base-sft-hh-harmless-4xh200-batch-64-20260417-214452](https://huggingface.co/jackf857/qwen3-8b-base-sft-hh-harmless-4xh200-batch-64-20260417-214452) on the Anthropic/hh-rlhf dataset.
It achieves the following results on the evaluation set:
- Loss: 0.5368
- Fcm Dpo/beta: 0.0530
- Margin Dpo/margin Mean: 11.7010
- Margin Dpo/margin Std: 18.7863
- Logps/chosen: -92.3002
- Logps/rejected: -113.7958
- Logps/ref Chosen: -86.9018
- Logps/ref Rejected: -96.6964
- Logits/chosen: 1.6311
- Logits/rejected: 1.4933

## Model description

More information needed

## Intended uses & limitations

More information needed

## Training and evaluation data

More information needed

## Training procedure

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 5e-07
- train_batch_size: 8
- eval_batch_size: 8
- seed: 42
- distributed_type: multi-GPU
- num_devices: 4
- gradient_accumulation_steps: 2
- total_train_batch_size: 64
- total_eval_batch_size: 32
- 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
- lr_scheduler_warmup_ratio: 0.1
- num_epochs: 1

### Training results

| Training Loss | Epoch  | Step | Validation Loss | Fcm Dpo/beta | Margin Dpo/margin Mean | Margin Dpo/margin Std | Logps/chosen | Logps/rejected | Logps/ref Chosen | Logps/ref Rejected | Logits/chosen | Logits/rejected |
|:-------------:|:------:|:----:|:---------------:|:------------:|:----------------------:|:---------------------:|:------------:|:--------------:|:----------------:|:------------------:|:-------------:|:---------------:|
| 1.3231        | 0.1512 | 100  | 0.6549          | 0.1000       | 0.8834                 | 1.9512                | -86.2078     | -96.8858       | -86.9018         | -96.6964           | 1.6085        | 1.4983          |
| 1.1388        | 0.3023 | 200  | 0.5426          | 0.2115       | 2.9544                 | 4.9177                | -81.7111     | -94.4601       | -86.9018         | -96.6964           | 1.6653        | 1.5454          |
| 1.1386        | 0.4535 | 300  | 0.5411          | 0.1198       | 5.2175                 | 8.5703                | -85.9098     | -100.9220      | -86.9018         | -96.6964           | 1.5470        | 1.4237          |
| 1.209         | 0.6047 | 400  | 0.5380          | 0.0890       | 6.8529                 | 11.1007               | -85.3711     | -102.0186      | -86.9018         | -96.6964           | 1.6431        | 1.5118          |
| 1.0608        | 0.7559 | 500  | 0.5388          | 0.0570       | 10.7654                | 17.4069               | -90.8454     | -111.4054      | -86.9018         | -96.6964           | 1.2932        | 1.1717          |
| 1.1399        | 0.9070 | 600  | 0.5368          | 0.0530       | 11.7010                | 18.7863               | -92.3002     | -113.7958      | -86.9018         | -96.6964           | 1.6311        | 1.4933          |


### Framework versions

- Transformers 4.51.0
- Pytorch 2.3.1+cu121
- Datasets 2.21.0
- Tokenizers 0.21.4