README.md

---
license: llama3
base_model: Magpie-Align/Llama-3-8B-Magpie-Align-SFT-v0.1
tags:
- alignment-handbook
- axolotl
- trl
- dpo
- sft
- generated_from_trainer
datasets:
- princeton-nlp/llama3-ultrafeedback
- Magpie-Align/Magpie-Pro-MT-300K-v0.1
model-index:
- name: Llama-3-8B-Magpie-Align-v0.1
  results: []
language:
- en
---

[![Magpie](magpie_logo.png)](https://huggingface.co/spaces/flydust/Chat-with-Magpie)
## 🔥 Chat with Magpie [Here](https://huggingface.co/spaces/flydust/Chat-with-Magpie)!

# 🐦 Llama-3-8B-Magpie-Align-v0.1

Project Web: [https://magpie-align.github.io/](https://magpie-align.github.io/)

Online Model Demo: [https://huggingface.co/spaces/flydust/Chat-with-Magpie](https://huggingface.co/spaces/flydust/Chat-with-Magpie)

Arxiv Technical Report: [https://arxiv.org/abs/2406.08464](https://arxiv.org/abs/2406.08464)

Codes: [https://github.com/magpie-align/magpie](https://github.com/magpie-align/magpie)

## Model Overview

This model is an aligned version of [meta-llama/Meta-Llama-3-8B](https://huggingface.co/meta-llama/Meta-Llama-3-8B). We apply the following pipeline:
- We first use [Magpie-Align/Magpie-Pro-MT-300K-v0.1](https://huggingface.co/datasets/Magpie-Align/Magpie-Pro-MT-300K-v0.1) dataset and perform SFT -> [Magpie-Align/Llama-3-8B-Magpie-Align-SFT-v0.1](https://huggingface.co/Magpie-Align/Llama-3-8B-Magpie-Align-SFT-v0.1)
- We then perform DPO on the [princeton-nlp/llama3-ultrafeedback](https://huggingface.co/datasets/princeton-nlp/llama3-ultrafeedback) dataset.

The overall performance is even better than the official Llama-3-8B-Instruct Model!

- **Alpaca Eval 2 (vs GPT-4-Turbo-1106): 38.52 (LC), 38.47 (WR)**
- **Alpaca Eval 2 (vs Llama-3-8B-Instruct): 69.37 (LC), 70.05 (WR)**
- **Arena Hard: 32.4**
- **WildBench: 39.3 ((was) Best <30B Model! 🏆)**
- **Zero-Eval GSM: 54.62**

## Model Performance

We compare our Llama-3-8B-Magpie-Align with official and other **open-aligned LLMs** that have been fine-tuned from base models and have publicly released their training datasets. The results are as follows:

```
+---------------------------------------------+--------------------+--------------------+-----------------------+------------+
|               Aligned Model ID              |      MT-Bench      |   Alpaca Eval 2    |     Alpaca Eval 2     | Arena Hard |
|                                             |                    | (GPT-4-Turbo-1106) | (Llama-3-8B-Instruct) |            |
+---------------------------------------------+------+------+------+----------+---------+-----------+-----------+------------+
|                                             |  R1  |  R2  |  AVG |   LC WR  |    WR   |   LC WR   |     WR    |    Score   |
+---------------------------------------------+------+------+------+----------+---------+-----------+-----------+------------+
| meta-llama/Meta-Llama-3-8B-Instruct         | 8.31 | 7.65 | 7.98 |   22.92  |  22.57  |     50    |     50    |    20.6    |
+---------------------------------------------+------+------+------+----------+---------+-----------+-----------+------------+
| princeton-nlp/Llama-3-Base-8B-SFT-DPO       | 8.12 | 7.23 | 7.67 |   17.71  |  15.34  |   43.73   |   38.80   |    14.8    |
+---------------------------------------------+------+------+------+----------+---------+-----------+-----------+------------+
| NousResearch/Hermes-2-Pro-Llama-3-8B        | 8.05 | 7.35 | 7.70 |   15.60  |  12.86  |   36.37   |   30.52   |    11.5    |
+---------------------------------------------+------+------+------+----------+---------+-----------+-----------+------------+
| allenai/llama-3-tulu-2-dpo-8b               | 7.71 | 7.15 | 7.43 |   14.89  |  14.80  |   35.43   |   35.42   |    11.7    |
+---------------------------------------------+------+------+------+----------+---------+-----------+-----------+------------+
| cognitivecomputations/dolphin-2.9-llama3-8b | 7.97 | 6.98 | 7.47 |   12.50  |   8.79  |   32.67   |   22.80   |     8.2    |
+---------------------------------------------+------+------+------+----------+---------+-----------+-----------+------------+
| openchat/openchat-3.6-8b-20240522           | 7.83 | 7.23 | 7.53 |   17.70  |  12.53  |   41.30   |   30.79   |     6.7    |
+---------------------------------------------+------+------+------+----------+---------+-----------+-----------+------------+
| Magpie-Align/Llama-3-8B-Magpie-Align-v0.1   | 8.01 | 7.63 | 7.82 |   38.52  |  38.47  |   69.37   |   70.05   |    32.4    |
+---------------------------------------------+------+------+------+----------+---------+-----------+-----------+------------+
| Magpie-Align/Llama-3-8B-Magpie-Align-v0.2   | 7.81 | 7.64 | 7.73 |   49.86  |  51.98  |   75.17   |   78.20   |    37.5    |
+---------------------------------------------+------+------+------+----------+---------+-----------+-----------+------------+

```

## 👀 Other Information

**License**: Please follow [Meta Llama 3 Community License](https://llama.meta.com/llama3/license).

**Conversation Template**: Please use Llama 3 **official chat template** for the best performance.

**How to use it?** Please check the official [Llama 3 repository](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct#how-to-use) for detailed instructions. Simply replace the original `model_id` with `Magpie-Align/Llama-3-8B-Magpie-Align-v0.1`.

The detailed training pipeline is as follows.

## Stage 1: Supervised Fine-tuning

We use [Axolotl](https://github.com/OpenAccess-AI-Collective/axolotl) for SFT.

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 2e-05
- train_batch_size: 1
- eval_batch_size: 1
- seed: 42
- distributed_type: multi-GPU
- num_devices: 4
- gradient_accumulation_steps: 8
- total_train_batch_size: 32
- total_eval_batch_size: 4
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: cosine
- lr_scheduler_warmup_steps: 100
- num_epochs: 2

### Training results

| Training Loss | Epoch  | Step | Validation Loss |
|:-------------:|:------:|:----:|:---------------:|
| 0.8807        | 0.0007 | 1    | 0.9001          |
| 0.5113        | 0.3337 | 464  | 0.5178          |
| 0.4668        | 0.6673 | 928  | 0.4792          |
| 0.4492        | 1.0010 | 1392 | 0.4582          |
| 0.3498        | 1.3205 | 1856 | 0.4575          |
| 0.3525        | 1.6542 | 2320 | 0.4555          |

### Framework versions

- Transformers 4.40.2
- Pytorch 2.3.0+cu121
- Datasets 2.19.1
- Tokenizers 0.19.1

[<img src="https://raw.githubusercontent.com/OpenAccess-AI-Collective/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/OpenAccess-AI-Collective/axolotl)
<details><summary>See axolotl config</summary>

axolotl version: `0.4.0`
```yaml

base_model: meta-llama/Meta-Llama-3-8B
model_type: LlamaForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: Magpie-Align/Magpie-Pro-MT-300K-v0.1
    type: sharegpt
    conversation: llama3
dataset_prepared_path: last_run_prepared
val_set_size: 0.001
output_dir: ./out_Llama-3-8B-Magpie-Pro-300K-MT

sequence_len: 8192
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: true

gradient_accumulation_steps: 8
micro_batch_size: 1
num_epochs: 2
optimizer: paged_adamw_8bit
lr_scheduler: cosine
learning_rate: 2e-5

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: false

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
early_stopping_patience:
resume_from_checkpoint:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_steps: 100
evals_per_epoch: 3
eval_table_size:
saves_per_epoch: 3
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
  pad_token: <|end_of_text|>

```

</details><be>

## Stage 2: Direct Preference Optimization

We use [alignment handbook](https://github.com/huggingface/alignment-handbook) for DPO.

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 1e-06
- train_batch_size: 2
- eval_batch_size: 4
- seed: 42
- distributed_type: multi-GPU
- num_devices: 4
- gradient_accumulation_steps: 16
- total_train_batch_size: 128
- total_eval_batch_size: 16
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: cosine
- lr_scheduler_warmup_ratio: 0.1
- num_epochs: 1

### Training results

| Training Loss | Epoch  | Step | Validation Loss | Rewards/chosen | Rewards/rejected | Rewards/accuracies | Rewards/margins | Logps/rejected | Logps/chosen | Logits/rejected | Logits/chosen |
|:-------------:|:------:|:----:|:---------------:|:--------------:|:----------------:|:------------------:|:---------------:|:--------------:|:------------:|:---------------:|:-------------:|
| 0.628         | 0.2138 | 100  | 0.6641          | -0.8806        | -1.0146          | 0.6240             | 0.1340          | -362.7133      | -343.6060    | -0.7539         | -0.7528       |
| 0.6935        | 0.4275 | 200  | 0.6352          | -1.3660        | -1.6311          | 0.6545             | 0.2651          | -424.3628      | -392.1437    | -0.6649         | -0.6629       |
| 0.6376        | 0.6413 | 300  | 0.6178          | -1.3533        | -1.6413          | 0.6748             | 0.2880          | -425.3859      | -390.8818    | -0.6753         | -0.6758       |
| 0.5888        | 0.8550 | 400  | 0.6088          | -1.6321        | -1.9785          | 0.6829             | 0.3464          | -459.1051      | -418.7560    | -0.6440         | -0.6435       |

It achieves the following results on the evaluation set:
- Loss: 0.6084
- Rewards/chosen: -1.6265
- Rewards/rejected: -1.9735
- Rewards/accuracies: 0.6809
- Rewards/margins: 0.3470
- Logps/rejected: -458.6070
- Logps/chosen: -418.2021
- Logits/rejected: -0.6447
- Logits/chosen: -0.6439

### Framework versions

- Transformers 4.41.2
- Pytorch 2.3.1+cu121
- Datasets 2.20.0
- Tokenizers 0.19.1

<details><summary>See alignment handbook config</summary>

```yaml

# Model arguments
model_name_or_path: Magpie-Align/Llama-3-8B-Magpie-Pro-MT-SFT-v0.1
torch_dtype: null

# Data training arguments
# For definitions, see: src/h4/training/config.py
dataset_mixer:
  princeton-nlp/llama3-ultrafeedback: 1.0
dataset_splits:
- train
- test
preprocessing_num_workers: 12

# DPOTrainer arguments
bf16: true
beta: 0.01
do_eval: true
evaluation_strategy: steps
eval_steps: 100
gradient_accumulation_steps: 16
gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: False
hub_model_id: Magpie-Align/Llama-3-8B-Magpie-Pro-MT-UltraDPO2
learning_rate: 1.0e-6
log_level: info
logging_steps: 1
lr_scheduler_type: cosine
max_length: 2048
max_prompt_length: 1800
num_train_epochs: 1
optim: adamw_torch
output_dir: data/magpie-pro-mt-ultradpo-1e-6
per_device_train_batch_size: 2
per_device_eval_batch_size: 4
push_to_hub: true
save_strategy: "steps"
save_steps: 100
save_total_limit: 1
seed: 42
warmup_ratio: 0.1

```

</details><be>

## Downstream Performance
| Datasets | Llama-3-8B-Magpie-Align-v0.1 |
| :--- | :---: |
| MMLU (5) | 64.61 | 
| ARC (25) | 62.03 |
| HellaSwag (25) | 82.10 |
| TruthfulQA (0) | 58.26 |
| Winogrande (5) | 73.01 |

## Paper Abstract

<details><summary>Click Here</summary>
High-quality instruction data is critical for aligning large language models (LLMs). Although some models, such as Llama-3-Instruct, have open weights, their alignment data remain private, which hinders the democratization of AI. High human labor costs and a limited, predefined scope for prompting prevent existing open-source data creation methods from scaling effectively, potentially limiting the diversity and quality of public alignment datasets. Is it possible to synthesize high-quality instruction data at scale by extracting it directly from an aligned LLM? We present a self-synthesis method for generating large-scale alignment data named Magpie. Our key observation is that aligned LLMs like Llama-3-Instruct can generate a user query when we input only the left-side templates up to the position reserved for user messages, thanks to their auto-regressive nature. We use this method to prompt Llama-3-Instruct and generate 4 million instructions along with their corresponding responses. We perform a comprehensive analysis of the extracted data and select 300K high-quality instances. To compare Magpie data with other public instruction datasets, we fine-tune Llama-3-8B-Base with each dataset and evaluate the performance of the fine-tuned models. Our results indicate that in some tasks, models fine-tuned with Magpie perform comparably to the official Llama-3-8B-Instruct, despite the latter being enhanced with 10 million data points through supervised fine-tuning (SFT) and subsequent feedback learning. We also show that using Magpie solely for SFT can surpass the performance of previous public datasets utilized for both SFT and preference optimization, such as direct preference optimization with UltraFeedback. This advantage is evident on alignment benchmarks such as AlpacaEval, ArenaHard, and WildBench.
</details><be>

## 📚 Citation

If you find the model, data, or code useful, please cite our paper:
```
@article{xu2024magpie,
	title={Magpie: Alignment Data Synthesis from Scratch by Prompting Aligned LLMs with Nothing}, 
	author={Zhangchen Xu and Fengqing Jiang and Luyao Niu and Yuntian Deng and Radha Poovendran and Yejin Choi and Bill Yuchen Lin},
	year={2024},
	eprint={2406.08464},
	archivePrefix={arXiv},
	primaryClass={cs.CL}
}
```

Please also cite the creators of preference datasets:

SimPO paper:
```
@article{meng2024simpo,
  title={{SimPO}: Simple preference optimization with a reference-free reward},
  author={Meng, Yu and Xia, Mengzhou and Chen, Danqi},
  journal={arXiv preprint arXiv:2405.14734},
  year={2024}
}
```

UltraFeedback paper:
```
@article{cui2023ultrafeedback,
  title={{UltraFeedback}: Boosting language models with high-quality feedback},
  author={Cui, Ganqu and Yuan, Lifan and Ding, Ning and Yao, Guanming and Zhu, Wei and Ni, Yuan and Xie, Guotong and Liu, Zhiyuan and Sun, Maosong},
  journal={arXiv preprint arXiv:2310.01377},
  year={2023}
}
```

ArmoRM paper:
```
@article{wang2024interpretable,
  title={Interpretable Preferences via Multi-Objective Reward Modeling and Mixture-of-Experts},
  author={Wang, Haoxiang and Xiong, Wei and Xie, Tengyang and Zhao, Han and Zhang, Tong},
  journal={arXiv preprint arXiv:2406.12845},
  year={2024}
}
```

**Questions?** Please contact [Zhangchen](https://zhangchenxu.com/) by email.