LongPO: Long Context Self-Evolution of Large Language Models through Short-to-Long Preference Optimization

This repo provides the official implementation of our paper "LongPO: Long Context Self-Evolution of Large Language Models through Short-to-Long Preference Optimization".

Updates

• [2024.2.17] 🚀 Release the code, data and checkpoints trained with LongPO.
• [2025.1.23] 🌟 LongPO has been accepted to ICLR 2025!

Highlights of LongPO

• Self-evolving long-context alignment without human/superior LLMs annotations.
• Extending context length while keeping aligned in one stage.
• No degradation on short-context capabilities.

Models and Training Data

Models	Base Model	Training Data	# Data Samples
Mistral-7B-LongPO-128K	Mistral-7B-Instruct-v0.2	HF Link	45K
Qwen2.5-7B-LongPO-128K	Qwen2.5-7B-Instruct	HF Link	32K
Mistral-7B-LongPO-256K-EXP*	Mistral-7B-LongPO-128K	HF Link	16K
Mistral-7B-LongPO-512K-EXP*	Mistral-7B-LongPO-128K	HF Link	2.5K

* indicates an experimental version (for rebuttal purposes) that may have not been fully tuned or provided with sufficient data to achieve convergence.

Training Process:

1. Prompt a short-context instruct LLM (e.g., Mistral-7B-Instruct-v0.2) to self-generate short-to-long preference data as illustrated in data_prepare.

2. Replace the (Flash) Attention module into Ulyssess (Flash) Attn using monkey patch to apply sequence parallel.

3. Using our custom LongPO Trainer: LongPOMTLMUlyssesTrainer

4. Train Script (using Mistral-7B-Instruct-v0.2 as example):

export training_length=131072
export gradient_accumulation_steps=8
export batch_size=1

accelerate launch \
--config_file playground/accelerate_single_node_zero3.yaml \
train/train_longpo.py \
    --model_name_or_path mistralai/Mistral-7B-Instruct-v0.2 \
    --ref_model_name_or_path mistralai/Mistral-7B-Instruct-v0.2 \
    --data_path /path/to/data \
    --bf16 True \
    --run_name mistral_longpo \
    --report_to wandb \
    --output_dir path/to/save \
    --num_train_epochs 1 \
    --per_device_train_batch_size $batch_size \
    --gradient_accumulation_steps $gradient_accumulation_steps \
    --save_strategy "steps" \
    --save_steps 500 \
    --evaluation_strategy "no" \
    --learning_rate 5e-7 \
    --weight_decay 0. \
    --warmup_ratio 0.1 \
    --lr_scheduler_type "cosine" \
    --optim "rmsprop" \
    --logging_steps 1 \
    --tf32 True \
    --model_max_length $training_length \
    --gradient_checkpointing True \
    --do_train True \
    --do_eval False \
    --do_predict False \
    --seed 42 \
    --use_sequence_parallel True \
    --dpo_beta 0.01 \
    --dpo_lambda 0.01 \
    --rope_theta 10000000

Evaluation

InfiniteBench

Model	Train/Claimed Length	En.Sum	En.QA	En.MC	AVG.
GPT-4-128K	128K	14.73	22.44	67.25	34.81
Qwen2-72B	128K	24.32ᵇ	7.03ᵇ	72.05ᵇ	34.47ᵇ
LLaMA 3.1-70B	128K	33.55ᵇ	36.08ᵇ	69.00ᵇ	46.21ᵇ
LLaMA 3.1-8B	128K	28.06ᵇ	30.47ᵇ	58.08ᵇ	38.87ᵇ
GLM-4-9B	128K	14.84ᵇ	9.51ᵇ	67.25ᵇ	30.53ᵇ
GLM-4-9B-1M	1M	28.3	9.7	68.6	35.53
LWM-7B-1M	1M	4.33ᵇ	0.0ᵇ	3.06ᵇ	2.46ᵇ
YaRN-Mistral-7B	128K	9.09	9.55	27.95	15.53
Mistral-7B	32K	22.13	4.93	14.41	13.82
- SFT	128K	23.44	13.45	53.21	30.03
- DPO	128K	15.21	10.34	48.14	25.56
- LongPO (iter1)	128K	27.05	23.51	67.25	39.27
- LongPO (iter2)	256K	28.16	24.43	66.35	39.65
- LongPO (iter3)	512K	29.10	27.85	66.67	41.21
Qwen2.5-7B	128K	22.89	6.08	52.4	27.12
- LongPO (iter1)	128K	32.06	17.32	72.05	40.48

• Our results are evaluated with greedy decoding.
• Baseline results marked with ᵇ are evaluated by us, while unmarked baseline results are sourced from their official report.

RULER

Model	NIAH	VT	AGG	QA	AVG (13 tasks)
Qwen2.5-7B-Instruct	82.10	80.09	74.50	54.30	76.50
Qwen2.5-7B-LongPO-128K	95.82	89.71	78.67	59.40	87.11
Mistral-7B-Instruct-v0.2	72.60	74.40	64.40	52.20	68.40
Mistral-7B-LongPO-128K	96.88	96.49	71.55	64.81	88.02
Mistral-7B-LongPO-256K-EXP	96.80	97.00	69.14	64.87	87.65
Mistral-7B-LongPO-512K-EXP	97.28	97.48	69.22	64.92	88.00

Short Context

Model	MMLU	ARC-C	Hellaswag	Winogrande	Avg
Mistral-7B-Instruct-v0.2	59.15	59.26	83.2	78.4	70.00
Mistral-7B-LongPO-128K	59.99	59.34	82.99	78.53	70.21
Mistral-7B-LongPO-256K-EXP	59.47	60.28	83.14	78.14	70.26
Mistral-7B-LongPO-512K-EXP	59.51	60.58	82.87	77.66	70.16
Qwen2.5-7B-Instruct	74.28	67.15	81.41	74.66	74.38
Qwen2.5-7B-LongPO-128K	73.64	65.70	80.82	74.98	73.79

Citation

If you find our project useful, hope you can star our repo and cite our paper as follows:

@inproceedings{
    chen2025longpo,
    title={Long{PO}: Long Context Self-Evolution of Large Language Models through Short-to-Long Preference Optimization},
    author={Guanzheng Chen and Xin Li and Michael Shieh and Lidong Bing},
    booktitle={The Thirteenth International Conference on Learning Representations},
    year={2025},
    url={https://openreview.net/forum?id=qTrEq31Shm}
}