URIAL: Untuned LLMs with Restyled In-context Alignment (ICLR'24: Rethinking Alignment via ICL)

This is part of the Rethinking Alignment (Re-Align) project by AI2 Mosaic.

📑 Paper: "The Unlocking Spell on Base LLMs: Rethinking Alignment via In-Context Learning" (ICLR 2024).

🛜 Website: https://allenai.github.io/re-align/.

🤗 Demo: BaseChat [https://huggingface.co/spaces/allenai/BaseChat_URIAL].

URIAL is a simple, tuning-free alignment method, URIAL (Untuned LLMs with Restyled In-context ALignment). URIAL achieves effective alignment purely through in-context learning (ICL), requiring as few as three constant stylistic examples and a system prompt. It's a strong baseline method for LLM alignment and shows comparable performance to fine-tuning based alignment. Apart from that, URIAL can also be used to study the science of LLMs, helping to understand alignment in a more controlled and interpretable manner.

Installation

conda create -n urial python=3.10  
conda activate urial
pip install vllm 
# conda create -p /net/nfs/mosaic/yuchenl/envs/urial python=3.10 
# conda activate /net/nfs/mosaic/yuchenl/envs/urial
pip install -r requirements.new.txt

URIAL Inference

An example script for running mistral (base) with urial prompts for alpaca_eval:

urial="inst_1k_v4" # urial prompt name -->  `urial_prompts/{urial}.txt`
output_dir="result_dirs/alpaca_eval/vllm_urial=${urial}/"  
CUDA_VISIBLE_DEVICES=0 python src/unified_infer.py \
    --urial $urial \
    --engine vllm \
    --model_name "mistralai/Mistral-7b-v0.1" \
    --tensor_parallel_size 1 \
    --dtype bfloat16 \
    --data_name "alpaca_eval" \
    --top_p 1.0 --temperature 0.3 --repetition_penalty 1.1 \
    --batch_size 16 --max_tokens 2048 \
    --output_folder $output_dir/

For more details, please refer to URIAL/src/unified_infer.py. Note that you can use the same method to run inference with aligned LLMs (by not setting --urial) too and also for other datasets. You could customize your own data/models in URIAL/src/unified_utils.py.

URIAL: ICL with constant prompts

🖼️ Click here to see a figure for the illustration of URIAL and other tuning-free Alignment methods.

Versions

As discussed here, a URIAL Prompt consists of K-shot stylistic in-context examples and a system prompt. The folder urial_prompts contains:

Suggested versions:

• inst_1k_v4.help (no refusal; K=3 examples; ~1k tokens)
• inst_1k_v4 (safer; K=3 examples; ~1k tokens)

Previous versions (used for the experiments in the arXiv version).

• URIAL-main (K=3; 1k tokens) -> inst_1k.txt
• URIAL-main (K=8; 2k tokens) -> inst_2k.txt
• URIAL-main (K=1; 0.5k tokens) -> inst_1shot.txt
• URIAL-ablation (K=3; 1k tokens) -> inst_1k_v2.txt
• URIAL-ablation (K=0; 0.15k tokens) -> inst_only.txt

Evaluation

AlpacaEval (fine-grained pairwise evaluation)

Show Tables

---

mistral-urial (#char=1105.7) VS Mistral-7B-Instruct-v0.1 (#char=1074.1) ⬇️

model	helpfulness	factuality	depth	engagement	clarity	safety
mistral-urial Win:	31.93	12.30	42.61	35.90	22.36	1.12
mistral-urial Tie:	38.88	73.04	19.63	31.68	60.62	98.39
mistral-urial Lose:	29.19	14.66	37.76	32.42	17.02	0.50

---

Llama-2-7b-urial (#char=1236.1) VS Llama-2-7b-chat-hf (#char=1455.7) ⬇️

model	helpfulness	factuality	depth	engagement	clarity	safety
Llama-2-7b-urial Win:	42.11	15.78	48.32	42.86	34.53	1.61
Llama-2-7b-urial Tie:	20.87	66.58	10.68	24.10	40.75	95.90
Llama-2-7b-urial Lose:	37.02	17.64	40.99	33.04	24.72	2.48

---

Llama-2-70b-urial (#char=1086.5) VS Llama-2-70b-chat-hf (#char=1524.0) ⬇️

model	helpfulness	factuality	depth	engagement	clarity	safety
Llama-2-70b-urial Win:	35.28	9.44	48.20	36.02	19.75	0.62
Llama-2-70b-urial Tie:	42.24	81.12	15.53	39.38	68.57	97.89
Llama-2-70b-urial Lose:	22.48	9.44	36.27	24.60	11.68	1.49

Scripts for URIAL/Aligned inference: run_scripts/alpaca_eval

Evaluation:

• evaluate/README.md
• Template for evaluation

MT-Bench

URIAL-MT Bench Scores (base LLMs + same URIAL prompts)

How to run: run_scripts/mt-bench/README.md

model	Turn 1	Turn 2	Overall
openai/gpt-4	8.96	9.03	8.99
openai/gpt-3.5-turbo	8.07	7.81	7.94
Base LLM + URIAL (3-shot ICL) ⬇️	--------	--------	---------
meta-llama/Llama-2-70b-hf	7.61	6.61	7.11
mistralai/Mixtral-8x7B-v0.1	7.69	6.19	6.94
mistralai/Mistral-7b-v0.1	7.49	5.86	6.67
01-ai/Yi-34B	7.19	6.16	6.67
google/gemma-7b	6.97	5.04	6.00
microsoft/phi-2 (2.7B)	7.04	4.66	5.85
meta-llama/Llama-2-13b-hf	6.27	4.41	5.34
01-ai/Yi-6B	5.96	3.99	4.97
meta-llama/Llama-2-7b-hf	5.75	3.91	4.83
google/gemma-2b	5.08	2.86	3.97
allenai/OLMo-7B	3.95	2.86	3.41

Just-Eval

Please find more details about our evaluation here: https://github.com/Re-Align/just-eval.

show more (the below content is outdated; will be updated soon)

Installation of Just-Eval

pip install git+https://github.com/Re-Align/just-eval.git
export OPENAI_API_KEY=<your secret key>

Reformatting output data

For example, if the output data is result_dirs/urial/inst_1k/Mistral-7B-v0.1.json, then run the following command to reformat the output data to result_dirs/urial/inst_1k/Mistral-7B-v0.1.to_eval.json.

python src/scripts/reformat.py result_dirs/urial/inst_1k/Mistral-7B-v0.1.json

Run Scoring

to_eval_file="result_dirs/urial/inst_1k/Mistral-7B-v0.1.to_eval.json"
run_name="Mistral-URIAL"
# GPT-4 for first five aspects on 0-800 examples 
just_eval \
    --mode "score_multi" \
    --model "gpt-4-0314" \
    --start_idx 0 \
    --end_idx 800 \
    --first_file $to_eval_file \
    --output_file "result_dirs/just-eval_results/${run_name}.score_multi.gpt-4.json"

# GPT-3.5-turbo for the safety aspect on 800-1000 examples
just_eval \
        --mode "score_safety" \
        --model "gpt-3.5-turbo-0613" \
        --first_file $to_eval_file \
        --start_idx 800 --end_idx 1000 \
        --output_file "result_dirs/just-eval_results/${run_name}.score_safety.chatgpt.json"  

Citation

@inproceedings{
    Lin2024ReAlign,
    title={The Unlocking Spell on Base LLMs: Rethinking Alignment via In-Context Learning},
    author={Bill Yuchen Lin and Abhilasha Ravichander and Ximing Lu and Nouha Dziri and Melanie Sclar and Khyathi Chandu and Chandra Bhagavatula and Yejin Choi},
    booktitle={International Conference on Learning Representations},
    year={2024},
    url={https://arxiv.org/abs/2312.01552}
}