Awesome

🐋 Humback

An unofficial implementation of Self-Alignment with Instruction Backtranslation .

The proposed Humback is a novel framework that can augment the instruction data for supervised fine-tuning with high quality.

🚧 Currently, this repo is under construction and not finished.

🌴 Dependencies

• Python==3.11.4
• PyTorch==2.0.1
• Others: requirements.txt

🚀 QuickStart

Procedure (2 iters):

1. Prepare seed data and unlabelled data.
2. Train the backward model $M_{yx}$ on the reversed seed data.
3. Self-augment the seed data via $M_{yx}$.
4. Train a forward model $M_{0}$ on the seed data.
5. Self-curate the unlabelled data $A_{k}^{(1)}$ via $M_{0}$ (tag quality scores).
6. Train a forward model $M_{1}$ on the self-curated unlabelled data $A_{k}^{(1)}$.
7. Use $M_{1}$ to self-curate the unlabelled data $A_{k}^{(2)}$.
8. Train a forward model $M_{2}$ on the self-curated unlabelled data $A_{k}^{(2)}$.

Seed Data Pre-processing

We follow the original paper and use oasst1 to construct the seed data.

The processed data could be found here .

$ bash data/seed/download.sh
$ python data/seed/convert.py
# #data: 3286, #dump: 3200
# Instruction len: 149±266, Response len: 1184±799

Unlabelled Data Pre-processing

Since ClueWeb22 is not a free open-source dataset, we sample texts from falcon-refinedweb instead.

The processed data could be found here .

$ python data/unlabelled/falcon_refinedweb.py

Train Backward Model $M_{yx}$

# The first Myx training takes about 30min (on the seed data)
$ bash scripts/train_backward_Myx.sh

The pre-trained $M_{yx}$ is available at Huggingface.

Self-Augmentation via $M_{yx}$

The augmentation data is available at Huggingface .

# Taking about 6:40:45 on the unlabelled data with 8*A100
$ bash scripts/self_aug.sh

Train Seed Model $M_{0}$

Hyper parameters are the same as $M_{yx}$.

$ bash scripts/train_seed.sh

The pre-trained $M_{0}$ is available at Huggingface (Uploading).

Self-Curation Prompting

The curated data is available at Huggingface .

# 33:54:45 with 8*A100 on 482,963 samples
$ bash scripts/self_curation.sh
# scores: [('None', 217203), ('4', 119211), ('3', 102756), ('5', 21301), ('1', 13083), ('2', 9288), ('8', 19), ('0', 15), ('9', 14), ('7', 11), ('6', 9), ('10', 4), ('91', 3), ('83', 2), ('20', 2), ('14', 2), ('75', 2), ('92', 2), ('72', 1), ('93', 1), ('28', 1), ('19', 1), ('728', 1), ('17', 1), ('16', 1), ('100', 1), ('237', 1), ('13', 1), ('73', 1), ('38', 1), ('87', 1), ('94', 1), ('98', 1), ('64', 1), ('52', 1), ('27', 1), ('24', 1), ('762', 1), ('266', 1), ('225', 1), ('80', 1), ('267', 1), ('99', 1), ('90', 1), ('63', 1), ('97', 1), ('78', 1), ('40', 1), ('1986', 1), ('47', 1), ('66', 1), ('45', 1), ('10502', 1), ('21', 1)]
# Number of qualified results (scores=5): 21301/482963
# instruction len: 198 ± 351
# response len: 1601 ± 345
# ---------------------------------------
# v2: (Strict Curation Score Matching: add `$` to the matching regex):
# Scores: [('None', 322324), ('3', 71851), ('4', 53120), ('5', 16460), ('1', 11921), ('2', 7260), ('0', 10), ('7', 4), ('6', 3), ('19', 1), ('8', 1), ('16', 1), ('13', 1), ('10', 1), ('23', 1), ('9', 1), ('90', 1), ('92', 1), ('45', 1)]
# Number of qualified results (scores=5): 15521/482963
# instruction len: 124 ± 113
# response len: 1611 ± 345
# ---------------------------------------
$ cat outputs/m1/unlabelled_curated_data.jsonl data/seed/seed.jsonl > data/curated/m1.jsonl

Train Models $M_{i}$

Most hyper parameters are the same as $M_{yx}$ except for the number of steps (the original Humback trains 1600 steps on 512k samples).

# change the `--data_path` in `scripts/train_seed.sh`
$ bash scripts/train_seed.sh

📑 Experimental Results

Other models: HuggingFaceH4/open_llm_leaderboard .

The results and the trend are not as good as the original paper, but the performance of $M_{0}$ is better than vanilla llama2-7b. Specifically, Humback $M_{1}$ is worse than $M_{0}$, and the different system prompts seem not be helpful on these benchmarks. By the way, although $M_{0}$ is good at these benchmarks, it may be not good at generating high quality and diversified responses on more tasks. Further experiments should be conducted to verify the effectiveness of the reproduced Humback $M_{0}$ (e.g. alpaca_eval with GPT4 as the judge).

Possible reasons are:

1. The backward model $M_{yx}$ is not good enough to generate high quality instructions.
2. The seed model $M_{0}$ is not competent to evaluate the generated quality (not all scores are ranging from 1 to 5).

Since I don't have GPT4 API keys, chatgpt_fn is used as the evaluator here (as introduced in alpaca_eval):

                       win_rate  standard_error  n_total  avg_length
gpt4                      73.79            1.54      805        1365
claude                    70.37            1.60      805        1082
chatgpt                   66.09            1.66      805         811
wizardlm-13b              65.16            1.67      805         985
vicuna-13b                64.10            1.69      805        1037
guanaco-65b               62.36            1.71      805        1249
oasst-rlhf-llama-33b      62.05            1.71      805        1079
alpaca-farm-ppo-human     60.25            1.72      805         803
falcon-40b-instruct       56.52            1.74      805         662
text_davinci_003          50.00            0.00      805         307
alpaca-7b                 45.22            1.74      805         396
HumbackM0                 32.30            1.65      805         548
text_davinci_001          28.07            1.56      805         296
HumbackM1                 23.35            1.49      805        1522

🔥 Further discussions are fully welcomed.

📝 TODO

• train more steps on $M_{i}$.
• remove system prompts when training $M_{0}$, $M_{i}$ and $M_{yx}$.

💌 Acknowledgments

• Paper: Self-Alignment with Instruction Backtranslation
• Code: FastChat
• Code: vLLM
• Code: stanford_alpaca
• Code: transformers

📜 Reference

@misc{li2023selfalignment,
    title={Self-Alignment with Instruction Backtranslation},
    author={Xian Li and Ping Yu and Chunting Zhou and Timo Schick and Luke Zettlemoyer and Omer Levy and Jason Weston and Mike Lewis},
    year={2023},
    eprint={2308.06259},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}