luq 0.0.2

Language Models Uncertainty Quantification (LUQ)

Get Started

Install LUQ:

pip install luq

Use LUQ model for UQ

import luq
from luq.models import MaxProbabilityEstimator

model_id = "meta-llama/Meta-Llama-3-8B-Instruct"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")
# Create text generation pipeline
generator = pipeline("text-generation", model=model, tokenizer=tokenizer)

# sample from LLM
samples = luq.llm.generate_n_samples_and_answer(
    pipeline,
    prompt="A, B, C, or D"
)

mp_estimator = MaxProbabilityEstimator()
print(mp_estimator.estimate_uncertainty(samples))

Uncertainty Quantification Methods

Generally the uncertainty quantification in LUQ sample multiple responses from an LLM and analyse the

Contributing

Use pre-commit

pip install pre-commit
pre-commit install

Pipeline for dataset creation

Step 1. Create a processed version of a dataset.

mkdir data/coqa
python scripts/process_datasets.py \
    --dataset=coqa \
    --output=data/coqa/processed.json

import json

data = json.load(open("data/coqa/processed.json", "r"))
new_data = {"train": data["train"][:2], "validation": data["validation"][:2]}
json.dump(new_data, open("data/coqa/processed_short.json", "w"))

Step 2. Generate answers from LLMs and augment the dataset with the dataset.

python scripts/add_generations_to_dataset.py \
    --input-file=./data/coqa/processed_short.json\
    --output-file=./data/coqa/processed_gen_short.json\

Step 3. Check accuracy of the answers given

python scripts/eval_accuracy.py \
    --input-file=data/coqa/processed_gen_short.json \
    --output-file=data/coqa/processed_gen_acc_short.json \
    --model-name=gpt2 \
    --model-type=huggingface

Step 4. Upload the dataset to HuggingFace

python scripts/upload_dataset.py \
    --path=data/coqa/processed_gen_acc_short.json \
    --repo-id your-username/dataset-name \
    --token your-huggingface-token

Datasets end-to-end

In order to generate a dataset:

python scripts/gen_dataset.py --input-file=./data/dummy_data/raw_dummy.json --output-file=./output.json

When a dataset is created we can augment it with accuracies checked by another LLM:

python scripts/eval_accuracy.py --input-file ./output.json --output-file test.json --model-name=gpt2 --model-type=huggingface