AlignTDS

Analyzing the Alignment of LLMs through the lens of Token Distribution Shift (TDS). Part of the Re-Align project by AI2 Mosaic. More info on our website and in our paper.

Alignment as Token Distribution Shifts 🔄

What changes does alignment tuning bring? 🧐

Analysis of TDS: Our approach involves comparing token distributions between base and aligned Large Language Models (LLMs) to understand the impact of alignment tuning.

The Analysis Pipeline ⚙️

1. Choose a pair of LLMs (e.g., Llama-2 and Llama-2-chat).
2. Get answer o from the aligned LLM.
3. Input the context to the base LLM and get the token distribution for the next position P_base.
4. Analyze the differences in distribution to understand the effects of alignment tuning.

Types of Token Positions Based on TDS 📊

• Unshifted positions: 🏠 Aligned token is also top 1 in P_base.
• Marginal positions: 🌿 Aligned token ranks 2nd or 3rd by P_base.
• Shifted positions: 🚀 Aligned token is outside the top 3 in P_base.

---

Web Demos for TDS analysis 🌐

• Visualize token distribution shifts easily with our web demos:
  • TDS demo: Llama-2-7b vs Llama-2-7b-chat (shifted ratio: 7.8%)
  • TDS demo: Llama-2-7b vs Vicuna-7b-v1.5 (shifted ratio: 4.8%)
  • TDS demo: Mistral-7b vs Mistral-7b-instruct (shifted ratio: 5.2%)

Key Findings 🔑

1. Only a small fraction of tokens are affected by alignment. The base and aligned LLMs usually share the same top-ranked tokens.
2. Alignment mainly changes stylistic elements, around 5-8% of positions.
3. Earlier tokens are more critical for alignment. The top token of the aligned model is often in the top 5 of the base model.
4. Base LLMs are already primed to follow instructions given an appropriate context.

Token Distribution Shift Analysis

1. Knowledge content comes from base LLMs.

Click to show/hide image 🖼️

2. TDS across different LLM pairs.

Click to show/hide images 🖼️

3. Learnings from alignment tuning.

Click to show/hide image 🖼️

4. TDS diminishes over time during decoding.

Click to show/hide images 🖼️

Procedures 🛠️

Generate outputs from aligned models

We use a generated output file containing the responses of aligned models. Filepath example: data/Llama-2-7b-chat-hf.json. See the repo URIAL for generation details.

Run Logit Analysis 📊

Save the token Logits of aligned models

# i2i   
instruct_data_file="data/Llama-2-7b-chat-hf.json"
logits_folder="saved_logits/just_eval_1000/llama2/shards/"
# i2i
mkdir -p $logits_folder 
n_shards=4 # or 1 if you only have one gpu
shard_size=250 # or 1000 if you only have one gpu
start_gpu=0
for ((start = 0, end = (($shard_size)), gpu = $start_gpu; gpu < $n_shards+$start_gpu; start += $shard_size, end += $shard_size, gpu++)); do
    CUDA_VISIBLE_DEVICES=$gpu python src/logit_analysis.py \
                --data_file $instruct_data_file \
                --logits_folder $logits_folder \
                --pair llama \
                --mode i2i \
                --start $start --end $end &  
done
# Merge the shards
python src/scripts/merge_logits.py saved_logits/just_eval_1000/llama/ llama i2i

Save the token logits of base models

logits_folder="saved_logits/just_eval_1000/llama2_tp/shards/"
mkdir -p $logits_folder
n_shards=4
shard_size=250
start_gpu=0
for ((start = 0, end = (($shard_size)), gpu = $start_gpu; gpu < $n_shards+$start_gpu; start += $shard_size, end += $shard_size, gpu++)); do
    CUDA_VISIBLE_DEVICES=$gpu python src/logit_analysis.py \
                --data_file $instruct_data_file \
                --enable_template \
                --logits_folder $logits_folder \
                --pair llama2 \
                --mode i2b \
                --i2i_pkl_file saved_logits/just_eval_1000/llama2/llama2-i2i.pkl \
                --start $start --end $end & 
done
# Merge the shards
python src/scripts/merge_logits.py saved_logits/just_eval_1000/llama2_tp/ llama2 i2b

Data Reformatting

python src/demo/data_prep.py llama2_tp saved_logits/just_eval_1000/llama2/llama2-i2i.pkl saved_logits/just_eval_1000/llama2_tp/llama2-i2b.pkl

Generate HTML pages for visualization

python src/demo/generate_html.py llama2_tp

TODOs 📝

• Integrate model generation into the logit computation process.
• Use vllm lib for efficiency improvements.
• Create an interactive demo.
• Add more data from larger LLMs.
• Compare models fine-tuned in different ways.

Citation 📄

@article{Lin2023ReAlign,
    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},
    journal = {ArXiv preprint},
    title = {The Unlocking Spell on Base LLMs: Rethinking Alignment via In-Context Learning},
    year = {2023}
}