π¦ MagpieLM-8B-Chat-v0.1
π§ About This Model
Model full name: Llama3.1-MagpieLM-8B-Chat-v0.1
This model is an aligned version of meta-llama/Meta-Llama-3.1-8B, which achieves state-of-the-art performance among open-aligned SLMs. It even outperforms larger open-weight models including Llama-3-8B-Instruct, Llama-3.1-8B-Instruct, Qwen-2-7B-Instruct, and Gemma-2-9B-it.
We apply the following standard alignment pipeline with two carefully crafted synthetic datasets.
We first perform SFT using Magpie-Align/MagpieLM-SFT-Data-v0.1.
- SFT Model Checkpoint: Magpie-Align/MagpieLM-8B-SFT-v0.1
We then perform DPO on the Magpie-Align/MagpieLM-DPO-Data-v0.1 dataset.
π₯ Benchmark Performance
Greedy Decoding
- Alpaca Eval 2: 58.18 (LC), 62.38 (WR)
- Arena Hard: 48.4
- WildBench WB Score (v2.0625): 44.72
Benchmark Performance Compare to Other SOTA SLMs
π Other Information
License: Please follow Meta Llama 3.1 Community License.
Conversation Template: Please use the Llama 3 chat template for the best performance.
Limitations: This model primarily understands and generates content in English. Its outputs may contain factual errors, logical inconsistencies, or reflect biases present in the training data. While the model aims to improve instruction-following and helpfulness, it isn't specifically designed for complex reasoning tasks, potentially leading to suboptimal performance in these areas. Additionally, the model may produce unsafe or inappropriate content, as no specific safety training were implemented during the alignment process.
π§ How to use it?
Please update transformers to the latest version by pip install git+https://github.com/huggingface/transformers
.
You can then run conversational inference using the Transformers pipeline
abstraction or by leveraging the Auto classes with the generate()
function.
import transformers
import torch
model_id = "MagpieLM-8B-Chat-v0.1"
pipeline = transformers.pipeline(
"text-generation",
model=model_id,
model_kwargs={"torch_dtype": torch.bfloat16},
device_map="auto",
)
messages = [
{"role": "system", "content": "You are Magpie, a friendly AI assistant."},
{"role": "user", "content": "Who are you?"},
]
outputs = pipeline(
messages,
max_new_tokens=256,
)
print(outputs[0]["generated_text"][-1])
Alignment Pipeline
The detailed alignment pipeline is as follows.
Stage 1: Supervised Fine-tuning
We use Axolotl for SFT. Please refer to the model card of SFT checkpoint and below for detailed configurations.
See axolotl config
axolotl version: 0.4.1
base_model: meta-llama/Meta-Llama-3.1-8B
model_type: LlamaForCausalLM
tokenizer_type: AutoTokenizer
chat_template: llama3
load_in_8bit: false
load_in_4bit: false
strict: false
main_process_port: 0
datasets:
- path: Magpie-Align/MagpieLM-SFT-Data-v0.1
type: sharegpt
conversation: llama3
dataset_prepared_path: last_run_prepared
val_set_size: 0.001
output_dir: axolotl_out/MagpieLM-8B-SFT-v0.1
sequence_len: 8192
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: true
wandb_project: SynDa
wandb_entity:
wandb_watch:
wandb_name: MagpieLM-8B-SFT-v0.1
wandb_log_model:
hub_model_id: Magpie-Align/MagpieLM-8B-SFT-v0.1
gradient_accumulation_steps: 32
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_ratio: 0.1
evals_per_epoch: 5
eval_table_size:
saves_per_epoch:
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
pad_token: <|end_of_text|>
Stage 2: Direct Preference Optimization
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 2e-07
- 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 |
---|---|---|---|---|---|---|---|---|---|---|---|
0.686 | 0.0653 | 100 | 0.6856 | -0.0491 | -0.0616 | 0.6480 | 0.0125 | -471.3315 | -478.8181 | -0.7034 | -0.7427 |
0.6218 | 0.1306 | 200 | 0.6277 | -0.6128 | -0.7720 | 0.6960 | 0.1591 | -542.3653 | -535.1920 | -0.7771 | -0.8125 |
0.5705 | 0.1959 | 300 | 0.5545 | -2.4738 | -3.0052 | 0.7270 | 0.5314 | -765.6894 | -721.2881 | -0.7894 | -0.8230 |
0.4606 | 0.2612 | 400 | 0.5081 | -2.6780 | -3.3782 | 0.7560 | 0.7002 | -802.9893 | -741.7116 | -0.6813 | -0.7247 |
0.4314 | 0.3266 | 500 | 0.4787 | -3.6697 | -4.6026 | 0.7630 | 0.9329 | -925.4283 | -840.8740 | -0.6189 | -0.6691 |
0.449 | 0.3919 | 600 | 0.4533 | -3.7414 | -4.8019 | 0.7820 | 1.0604 | -945.3563 | -848.0514 | -0.6157 | -0.6681 |
0.4538 | 0.4572 | 700 | 0.4350 | -4.3858 | -5.6549 | 0.7890 | 1.2690 | -1030.6561 | -912.4920 | -0.5789 | -0.6331 |
0.35 | 0.5225 | 800 | 0.4186 | -4.7129 | -6.1662 | 0.8010 | 1.4533 | -1081.7843 | -945.1964 | -0.5778 | -0.6347 |
0.4153 | 0.5878 | 900 | 0.4108 | -4.9836 | -6.5320 | 0.7970 | 1.5484 | -1118.3677 | -972.2631 | -0.5895 | -0.6474 |
0.3935 | 0.6531 | 1000 | 0.3999 | -4.4303 | -5.9370 | 0.8110 | 1.5067 | -1058.8646 | -916.9379 | -0.6016 | -0.6598 |
0.3205 | 0.7184 | 1100 | 0.3950 | -5.1884 | -6.8827 | 0.8010 | 1.6943 | -1153.4371 | -992.7452 | -0.5846 | -0.6452 |
0.3612 | 0.7837 | 1200 | 0.3901 | -5.0426 | -6.7179 | 0.8040 | 1.6753 | -1136.9619 | -978.1701 | -0.6046 | -0.6637 |
0.3058 | 0.8490 | 1300 | 0.3877 | -5.1224 | -6.8428 | 0.8040 | 1.7204 | -1149.4465 | -986.1475 | -0.6087 | -0.6690 |
0.3467 | 0.9144 | 1400 | 0.3871 | -5.2335 | -6.9809 | 0.8090 | 1.7474 | -1163.2629 | -997.2610 | -0.6071 | -0.6672 |
0.3197 | 0.9797 | 1500 | 0.3867 | -5.1502 | -6.8793 | 0.8080 | 1.7291 | -1153.0979 | -988.9237 | -0.6120 | -0.6722 |
Framework versions
- Transformers 4.44.2
- Pytorch 2.4.1+cu121
- Datasets 3.0.0
- Tokenizers 0.19.1
See alignment handbook configs
# Customized Configs
model_name_or_path: Magpie-Align/MagpieLM-8B-SFT-v0.1
hub_model_id: Magpie-Align/MagpieLM-8B-Chat-v0.1
output_dir: alignment_handbook_out/MagpieLM-8B-Chat-v0.1
run_name: MagpieLM-8B-Chat-v0.1
dataset_mixer:
Magpie-Align/MagpieLM-DPO-Data-v0.1: 1.0
dataset_splits:
- train
- test
preprocessing_num_workers: 24
# DPOTrainer arguments
bf16: true
beta: 0.01
learning_rate: 2.0e-7
gradient_accumulation_steps: 16
per_device_train_batch_size: 2
per_device_eval_batch_size: 4
num_train_epochs: 1
max_length: 2048
max_prompt_length: 1800
warmup_ratio: 0.1
logging_steps: 1
lr_scheduler_type: cosine
optim: adamw_torch
torch_dtype: null
# use_flash_attention_2: true
do_eval: true
evaluation_strategy: steps
eval_steps: 100
gradient_checkpointing: true
gradient_checkpointing_kwargs:
use_reentrant: False
log_level: info
push_to_hub: true
save_total_limit: 0
seed: 42
report_to:
- wandb
π Citation
If you find the model, data, or code useful, please cite:
@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}
}
Contact
Questions? Contact:
- Zhangchen Xu [zxu9 at uw dot edu], and
- Bill Yuchen Lin [yuchenlin1995 at gmail dot com]
- Downloads last month
- 2,741
Model tree for Magpie-Align/MagpieLM-8B-Chat-v0.1
Base model
meta-llama/Llama-3.1-8B