FActScore

This is the official release accompanying our EMNLP 2023 paper, FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation. FActScore is available as a PIP package as well.

If you find FActScore useful, please cite:

@inproceedings{ factscore,
    title={ {FActScore}: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation },
    author={ Min, Sewon and Krishna, Kalpesh and Lyu, Xinxi and Lewis, Mike and Yih, Wen-tau and Koh, Pang Wei and Iyyer, Mohit and Zettlemoyer, Luke and Hajishirzi, Hannaneh },
    year={ 2023 },
    booktitle = { EMNLP },
    url={ https://arxiv.org/abs/2305.14251 }
}

Announcement

• 11/04/2023: The data we release includes human annotations of factual precision reported in Section 3 of the paper. If you want to download these human annotated data only, without other data, you can download it directly from this Google Drive link. We are also releasing FActScore results of 12 different LMs reported in Section 4.3 of the paper, in case you want to obtain them without running the code. Please refer to here.

Install

Make a new Python 3.7+ environment using virtualenv or conda.

pip install --upgrade factscore
python -m spacy download en_core_web_sm

Download the data

python -m factscore.download_data --llama_7B_HF_path "llama-7B"

This command does the following.

1. Download the knowledge source and example data.
2. Take the LLAMA 7B model and reconstruct Inst-LLAMA. This requires having access to HuggingFace weights of the LLAMA-7B model, which are added to the --llama_7B_HF_path flag. Follow this guide in order to obtain those weights. Skip the --llama_7B_HF_path if you would only like to use the ChatGPT version of FActScore.

Optional flags:

• --data_dir: directory to store the knowledge source and example data. .cache/factscore by default.
• --model_dir: directory to store Inst-LLAMA weights. .cache/factscore by default.

Troubleshooting:

• If you get a ERROR 429: Too Many Requests error while downloading the DB file, please download the DB from this Google Drive link and place it under --data_dir (.cache/factscore by default).
• If everything else fails, consider downloading the files manually from this link and placing them in --data_dir and --model_dir, see factscore/download_data.py for more details.

Running FActScore using a command line

We expect running FActScore costs about $1 of the API cost per 100 sentences. For instance, if you have 100 generations, each with 5 sentences on average, it costs $5 in total.

python -m factscore.factscorer --input_path {input_path} --model_name {estimator_name} --openai_key {openai_key}

• --input_path can be something like data/unlabeled/InstructGPT.jsonl. It should be a .jsonl format where each line contains topic (a topic entity that corresponds to the Wikipedia title) and output (a generation from the model).
• --model_name: retrieval+ChatGPT and retrieval+llama+npm (You can also use retrieval+ChatGPT+npm or retrieval+llama but we recommend the former two.)
• --openai_key: File containing OpenAI API Key.

Optional flags:

• --data_dir: Directory containing knowledge source, etc. .cache/factscore by default.
• --model_dir: Directory containing Inst-LLAMA weights. Skip if your model_name doesn't include llama. .cache/factscore by default.
• --cache_dir: Directory containing cache from API/models. .cache/factscore by default.
• --use_atomic_facts: If specified, it uses model-generated atomic facts released as part of our data instead of running the atomic fact generator. This will allow reproducing our results with no (or little if it still uses ChatGPT) cost. You can't specify it if you are running new model generations.
• --gamma: A hyperparameter for length penalty. 10 by default. It penalizes the score if the number of facts is less than gamma. 10 roughly corresponds to 2 sentences, so would penalize if the generation has less than 2 sentences. Usually, this would not change the ranking between systems unless some systems generate overly short responses all the time (e.g., models trained on NLP datasets without long-form generation tasks may do so). If you would like to turn off the length penalty completely, specify --gamma 0.
• --n_samples: If specified, it runs the model on a subset of the data.
• --verbose: If specified, it shows the progress bar.
• --print_rate_limit_error: It specified, it prints out rate limit errors from OpenAI API.
• --cost_estimate: This flag decides the type of OpenAI API cost estimation that we provide before calling it. It can be "consider_cache" (default) or "ignore_cache".
• --abstain_detection: This flag optionally enables automatic detection of abstained responses. By default this is disabled, but it is recommended to add your own function tailored to your model. The currently supported detectors are "generic" and "perplexity_ai", and their implementations can be found in factscore/abstain_detection.py. There are two methods to add your own abstain function: a) clone our GitHub repository to install factscore locally (pip install --editable .), and then add your function to factscore/abstain_detection.py directly; b) process your abstain detection outside our package, and use empty strings in the output key for the JSONL file used in --input_path.
• --knowledge_source: In case the default knowledge source (Wikipedia - 2023/04/01) will not be used, preprocess it using the instructions below, and then specify the knowledge_source name under this flag.

To evaluate your own LM

There're two sets of prompt entities, data/labeled/prompt_entities.txt (183 entities) and data/unlabeled/prompt_entities.txt (500 entities). Each line contains the name of the person (which is also a corresponding Wikipedia title). You can use the labeled version if you want to be compatible with the data under data/labeled (Section 3 and Section 4.2 in the paper), and use the unlabeled version if you want to be compatible with the data under data/unlabeled (Section 4.3 in the paper).

You can prompt your LM with your own prompt (we used Question: Tell me a bio of <entity>.) and use the following code.

from factscore.factscorer import FactScorer

fs = FactScorer(openai_key="...")

# topics: list of strings (human entities used to generate bios)
# generations: list of strings (model generations)
out = fs.get_score(topics, generations, gamma=10)
print (out["score"]) # FActScore
print (out["init_score"]) # FActScore w/o length penalty
print (out["respond_ratio"]) # % of responding (not abstaining from answering)
print (out["num_facts_per_response"]) # average number of atomic facts per response

Alternatively, you can create a .jsonl file, where each line has topic (entity name, exactly same as the one from .txt file) and output (generation from LM), and then use a command line above.

We recommend using (A) FactScorer(model_name="retrieval+ChatGPT") (default) or (B) FactScorer(model_name="retrieval+llama+npm"). They have 0.99 Pearson correlation. Here're results of a range of models, which you can easily reproduce through these command lines.

Model	% respond	# facts	FActScore from (A)	FActScore from (B)
GPT-4	88.2	60.8	73.1	59.9
ChatGPT	84.2	37.0	71.6	60.4
Alpaca 65B	100.0	17.1	55.6	46.3
InstructGPT	99.8	27.7	52.8	41.7
Alpaca 13B	100.0	16.6	47.7	40.3
Vicuna 13B	76.6	50.9	46.6	40.7
Alpaca 7B	100.0	17.4	39.7	36.5
Vicuna 7B	91.0	45.6	38.9	36.9
MPT Chat 7B	88.8	37.3	30.1	27.9
Oasst Pythia 12B	100.0	39.7	25.1	20.8
Dolly 12B	100.0	24.6	21.7	17.1
StableLM tuned 7B	66.6	38.0	17.3	16.3

% respond (% of responding instead of abstaining from answering) and # facts (# of atomic facts per valid response) indicate "factual recall" (how many pieces of information the model gives) and FActScore indicates "factual precision" (how accurate each piece of information the model gives is).

To use a custom knowledge source

By default, FActScore uses Wikipedia dump from 2023/04/01. But you can also use your own knowledge source!

The knolwedge source should be ready in a .jsonl format, where each line is a dictionary containing title and text. text can either be a string or a list of strings (e.g., sections).

from factscore.factscorer import FactScorer

fs = FactScorer()

# this will create a database using your file
# for English Wikipedia (18GB)), it takes ~8 hours
# once DB file is created, you can reuse it by only specifying `db_path`
fs.register_knowledge_source(name_of_your_knowledge_source,
                             data_path=path_to_jsonl_file,
                             db_path=path_to_output_db_file)

# now, when you compute a score, specify knowledge source to use
out = fs.get_score(topics, generations, knowledge_source=name_of_your_knowledge_source)
print (out["score"]) # FActScore
print (out["respond_ratio"]) # % of responding (not abstaining from answering)
print (out["num_facts_per_response"]) # average number of atomic facts per response

To see an example of constructing the ACL anthology knowledge source, see preprocessing/preprocess_acl.py.

FActScore results of the unlabeled data

You can easily reproduce FActScore results of 12 different LMs reported in Section 4.3 of the paper using this code. However, if you would like to obtain their predictions without running the code, you can download it from this Google Drive link.

Each file corresponds to the subject LM (LM that generates responses that we are validating). Each line is a dictionary:

• prompt: the initial prompt fed into the LM
• facts: atomic facts decomposed by the model
• LLAMA+NP_labels: labels to facts, verified by LLAMA+NP
• ChatGPT_labels: labels to facts, verified by ChatGPT

Note that the number of lines may be less than 500, because it excludes the cases where the model abstains from responding (e.g., it says "I don't know"). You can do # of lines / 500 to calculate the response ratio.

If you unzip the data and run the following code for verification, you will be able to get statistics that exactly match the statistics reported in the paper (Table 5 and Figure 3).

dirname = "factscore-unlabeled-predictions"
for fn in os.listdir(dirname):
    chatgpt_fs = []
    llama_fs = []
    n_facts = []
    with open(os.path.join(dirname, fn)) as f:
        for line in f:
            dp = json.loads(line)
            n_facts.append(len(dp["facts"]))
            if "ChatGPT_Labels" in dp:
                chatgpt_fs.append(np.mean([l=="S" for l in dp["ChatGPT_Labels"]]))
            llama_fs.append(np.mean([l=="S" for l in dp["LLAMA+NP_Labels"]]))
    print ("Model=%s\t(%.1f%% responding, %.1f facts/response)\tFactScore=%.1f (ChatGPT)\t%.1f (LLAMA)" % (
        fn.split(".")[0], len(n_facts)*100/500, np.mean(n_facts), np.mean(chatgpt_fs)*100, np.mean(llama_fs)*100
    ))