answer equivalence tiny Bert

README.md

@@ -1,3 +1,263 @@
----
-license: apache-2.0
----
+---
+inference: false
+license: mit
+language:
+- en
+metrics:
+- exact_match
+- f1
+- bertscore
+pipeline_tag: text-classification
+---
+# QA-Evaluation-Metrics
+
+[![PyPI version qa-metrics](https://img.shields.io/pypi/v/qa-metrics.svg)](https://pypi.org/project/qa-metrics/) 
+[![Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/17b7vrZqH0Yun2AJaOXydYZxr3cw20Ga6?usp=sharing)
+
+QA-Evaluation-Metrics is a fast and lightweight Python package for evaluating question-answering models and prompting of black-box and open-source large language models. It provides various basic and efficient metrics to assess the performance of QA models. 
+
+### Updates
+- Uopdated to version 0.2.8 
+  - Supports prompting OPENAI GPT-series models and Claude Series models now. (Assuimg OPENAI version > 1.0)
+  - Supports prompting various open source models such as LLaMA-2-70B-chat, LLaVA-1.5 etc by calling API from [deepinfra](https://deepinfra.com/models).
+
+
+## Installation
+* Python version >= 3.6
+* openai version >= 1.0
+
+
+To install the package, run the following command:
+
+```bash
+pip install qa-metrics
+```
+
+## Usage/Logistics
+
+The python package currently provides six QA evaluation methods. 
+- Given a set of gold answers, a candidate answer to be evaluated, and a question (if applicable), the evaluation returns True if the candidate answer matches any one of the gold answer, False otherwise.
+- Different evaluation methods have distinct strictness of evaluating the correctness of a candidate answer. Some have higher correlation with human judgments than others.
+- Normalized Exact Match and Question/Answer type Evaluation are the most efficient method. They are suitable for short-form QA datasets such as NQ-OPEN, Hotpot QA, TriviaQA, SQuAD, etc.
+- Question/Answer Type Evaluation and Transformer Neural evaluations are cost free and suitable for short-form and longer-form QA datasets. They have higher correlation with human judgments than exact match and F1 score when the length of the gold and candidate answers become long.
+- Black-box LLM evaluations are closest to human evaluations, and they are not cost-free.
+
+## Normalized Exact Match
+#### `em_match`
+
+Returns a boolean indicating whether there are any exact normalized matches between gold and candidate answers.
+
+**Parameters**
+
+- `reference_answer` (list of str): A list of gold (correct) answers to the question.
+- `candidate_answer` (str): The answer provided by a candidate that needs to be evaluated.
+
+**Returns**
+
+- `boolean`: A boolean True/False signifying matches between reference or candidate answers.
+
+```python
+from qa_metrics.em import em_match
+
+reference_answer = ["The Frog Prince", "The Princess and the Frog"]
+candidate_answer = "The movie \"The Princess and the Frog\" is loosely based off the Brother Grimm's \"Iron Henry\""
+match_result = em_match(reference_answer, candidate_answer)
+print("Exact Match: ", match_result)
+'''
+Exact Match:  False
+'''
+```
+
+## F1 Score
+#### `f1_score_with_precision_recall`
+
+Calculates F1 score, precision, and recall between a reference and a candidate answer.
+
+**Parameters**
+
+- `reference_answer` (str): A gold (correct) answers to the question.
+- `candidate_answer` (str): The answer provided by a candidate that needs to be evaluated.
+
+**Returns**
+
+- `dictionary`: A dictionary containing the F1 score, precision, and recall between a gold and candidate answer.
+
+```python
+from qa_metrics.f1 import f1_match,f1_score_with_precision_recall
+
+f1_stats = f1_score_with_precision_recall(reference_answer[0], candidate_answer)
+print("F1 stats: ", f1_stats)
+'''
+F1 stats:  {'f1': 0.25, 'precision': 0.6666666666666666, 'recall': 0.15384615384615385}
+'''
+
+match_result = f1_match(reference_answer, candidate_answer, threshold=0.5)
+print("F1 Match: ", match_result)
+'''
+F1 Match:  False
+'''
+```
+
+## Efficient and Robust Question/Answer Type Evaluation
+#### 1. `get_highest_score`
+
+Returns the gold answer and candidate answer pair that has the highest matching score. This function is useful for evaluating the closest match to a given candidate response based on a list of reference answers.
+
+**Parameters**
+
+- `reference_answer` (list of str): A list of gold (correct) answers to the question.
+- `candidate_answer` (str): The answer provided by a candidate that needs to be evaluated.
+- `question` (str): The question for which the answers are being evaluated.
+
+**Returns**
+
+- `dictionary`: A dictionary containing the gold answer and candidate answer that have the highest matching score.
+
+#### 2. `get_scores`
+
+Returns all the gold answer and candidate answer pairs' matching scores.
+
+**Parameters**
+
+- `reference_answer` (list of str): A list of gold (correct) answers to the question.
+- `candidate_answer` (str): The answer provided by a candidate that needs to be evaluated.
+- `question` (str): The question for which the answers are being evaluated.
+
+**Returns**
+
+- `dictionary`: A dictionary containing gold answers and the candidate answer's matching score.
+
+#### 3. `evaluate`
+
+Returns True if the candidate answer is a match of any of the gold answers.
+
+**Parameters**
+
+- `reference_answer` (list of str): A list of gold (correct) answers to the question.
+- `candidate_answer` (str): The answer provided by a candidate that needs to be evaluated.
+- `question` (str): The question for which the answers are being evaluated.
+
+**Returns**
+
+- `boolean`: A boolean True/False signifying matches between reference or candidate answers.
+
+
+```python
+from qa_metrics.pedant import PEDANT
+
+question = "Which movie is loosley based off the Brother Grimm's Iron Henry?"
+pedant = PEDANT()
+scores = pedant.get_scores(reference_answer, candidate_answer, question)
+max_pair, highest_scores = pedant.get_highest_score(reference_answer, candidate_answer, question)
+match_result = pedant.evaluate(reference_answer, candidate_answer, question)
+print("Max Pair: %s; Highest Score: %s" % (max_pair, highest_scores))
+print("Score: %s; PANDA Match: %s" % (scores, match_result))
+'''
+Max Pair: ('the princess and the frog', 'The movie "The Princess and the Frog" is loosely based off the Brother Grimm\'s "Iron Henry"'); Highest Score: 0.854451712151719
+Score: {'the frog prince': {'The movie "The Princess and the Frog" is loosely based off the Brother Grimm\'s "Iron Henry"': 0.7131625951317375}, 'the princess and the frog': {'The movie "The Princess and the Frog" is loosely based off the Brother Grimm\'s "Iron Henry"': 0.854451712151719}}; PANDA Match: True
+'''
+```
+
+```python
+print(pedant.get_score(reference_answer[1], candidate_answer, question))
+'''
+0.7122460127464126
+'''
+```
+
+## Transformer Neural Evaluation
+Our fine-tuned BERT model is on 🤗 [Huggingface](https://huggingface.co/Zongxia/answer_equivalence_bert?text=The+goal+of+life+is+%5BMASK%5D.). Our Package also supports downloading and matching directly. [distilroberta](https://huggingface.co/Zongxia/answer_equivalence_distilroberta), [distilbert](https://huggingface.co/Zongxia/answer_equivalence_distilbert), [roberta](https://huggingface.co/Zongxia/answer_equivalence_roberta), and [roberta-large](https://huggingface.co/Zongxia/answer_equivalence_roberta-large) are also supported now! 🔥🔥🔥
+
+#### `transformer_match`
+
+Returns True if the candidate answer is a match of any of the gold answers.
+
+**Parameters**
+
+- `reference_answer` (list of str): A list of gold (correct) answers to the question.
+- `candidate_answer` (str): The answer provided by a candidate that needs to be evaluated.
+- `question` (str): The question for which the answers are being evaluated.
+
+**Returns**
+
+- `boolean`: A boolean True/False signifying matches between reference or candidate answers.
+
+```python
+from qa_metrics.transformerMatcher import TransformerMatcher
+
+question = "Which movie is loosley based off the Brother Grimm's Iron Henry?"
+# Supported models: roberta-large, roberta, bert, distilbert, distilroberta
+tm = TransformerMatcher("roberta-large")
+scores = tm.get_scores(reference_answer, candidate_answer, question)
+match_result = tm.transformer_match(reference_answer, candidate_answer, question)
+print("Score: %s; bert Match: %s" % (scores, match_result))
+'''
+Score: {'The Frog Prince': {'The movie "The Princess and the Frog" is loosely based off the Brother Grimm\'s "Iron Henry"': 0.6934309}, 'The Princess and the Frog': {'The movie "The Princess and the Frog" is loosely based off the Brother Grimm\'s "Iron Henry"': 0.7400551}}; TM Match: True
+'''
+```
+
+## Prompting LLM For Evaluation
+
+Note: The prompting function can be used for any prompting purposes.
+
+###### OpenAI
+```python
+from qa_metrics.prompt_llm import CloseLLM
+model = CloseLLM()
+model.set_openai_api_key(YOUR_OPENAI_KEY)
+prompt = 'question: What is the Capital of France?\nreference: Paris\ncandidate: The capital is Paris\nIs the candidate answer correct based on the question and reference answer? Please only output correct or incorrect.'
+model.prompt_gpt(prompt=prompt, model_engine='gpt-3.5-turbo', temperature=0.1, max_tokens=10)
+
+'''
+'correct'
+'''
+```
+
+###### Anthropic
+```python
+model = CloseLLM()
+model.set_anthropic_api_key(YOUR_Anthropic_KEY)
+model.prompt_claude(prompt=prompt, model_engine='claude-v1', anthropic_version="2023-06-01", max_tokens_to_sample=100, temperature=0.7)
+
+'''
+'correct'
+'''
+```
+
+###### deepinfra (See below for descriptions of more models)
+```python
+from qa_metrics.prompt_open_llm import OpenLLM
+model = OpenLLM()
+model.set_deepinfra_key(YOUR_DEEPINFRA_KEY)
+model.prompt(message=prompt, model_engine='mistralai/Mixtral-8x7B-Instruct-v0.1', temperature=0.1, max_tokens=10)
+
+'''
+'correct'
+'''
+```
+
+If you find this repo avialable, please cite our paper:
+```bibtex
+@misc{li2024panda,
+      title={PANDA (Pedantic ANswer-correctness Determination and Adjudication):Improving Automatic Evaluation for Question Answering and Text Generation}, 
+      author={Zongxia Li and Ishani Mondal and Yijun Liang and Huy Nghiem and Jordan Lee Boyd-Graber},
+      year={2024},
+      eprint={2402.11161},
+      archivePrefix={arXiv},
+      primaryClass={cs.CL}
+}
+```
+
+
+## Updates
+- [01/24/24] 🔥 The full paper is uploaded and can be accessed [here](https://arxiv.org/abs/2402.11161). The dataset is expanded and leaderboard is updated.
+- Our Training Dataset is adapted and augmented from [Bulian et al](https://github.com/google-research-datasets/answer-equivalence-dataset). Our [dataset repo](https://github.com/zli12321/Answer_Equivalence_Dataset.git) includes the augmented training set and QA evaluation testing sets discussed in our paper. 
+- Now our model supports [distilroberta](https://huggingface.co/Zongxia/answer_equivalence_distilroberta), [distilbert](https://huggingface.co/Zongxia/answer_equivalence_distilbert), a smaller and more robust matching model than Bert!
+
+## License
+
+This project is licensed under the [MIT License](LICENSE.md) - see the LICENSE file for details.
+
+## Contact
+
+For any additional questions or comments, please contact [zli12321@umd.edu].

config.json

@@ -0,0 +1,26 @@
+{
+  "_name_or_path": "/srv/www/active-topic-modeling/ae_tune/models--google--bert_uncased_L-2_H-128_A-2/snapshots/30b0a37ccaaa32f332884b96992754e246e48c5f",
+  "architectures": [
+    "BertForSequenceClassification"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "classifier_dropout": null,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 128,
+  "initializer_range": 0.02,
+  "intermediate_size": 512,
+  "layer_norm_eps": 1e-12,
+  "max_position_embeddings": 512,
+  "model_type": "bert",
+  "num_attention_heads": 2,
+  "num_hidden_layers": 2,
+  "pad_token_id": 0,
+  "position_embedding_type": "absolute",
+  "problem_type": "single_label_classification",
+  "torch_dtype": "float32",
+  "transformers_version": "4.37.2",
+  "type_vocab_size": 2,
+  "use_cache": true,
+  "vocab_size": 30522
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3e3fcca05ef67be24f3ab588c63ac2f46cb6da5d84751fd0e8c6a25a246c4016
+size 133

special_tokens_map.json

@@ -0,0 +1,7 @@
+{
+  "cls_token": "[CLS]",
+  "mask_token": "[MASK]",
+  "pad_token": "[PAD]",
+  "sep_token": "[SEP]",
+  "unk_token": "[UNK]"
+}

tokenizer_config.json

@@ -0,0 +1,57 @@
+{
+  "added_tokens_decoder": {
+    "0": {
+      "content": "[PAD]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "100": {
+      "content": "[UNK]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "101": {
+      "content": "[CLS]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "102": {
+      "content": "[SEP]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "103": {
+      "content": "[MASK]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "clean_up_tokenization_spaces": true,
+  "cls_token": "[CLS]",
+  "do_basic_tokenize": true,
+  "do_lower_case": true,
+  "mask_token": "[MASK]",
+  "model_max_length": 1000000000000000019884624838656,
+  "never_split": null,
+  "pad_token": "[PAD]",
+  "sep_token": "[SEP]",
+  "strip_accents": null,
+  "tokenize_chinese_chars": true,
+  "tokenizer_class": "BertTokenizer",
+  "unk_token": "[UNK]"
+}