Sentence Transformer for Audit Retrieval Question-Answering (STAR-QA)

Sentence Transformer for Audit Retrieval Question-Answering (STAR-QA) is a fine-tuned sentence-transformers model based on ALL-MPNET-BASE-V2. It has been developed to produce high-performance embeddings for audit, risk-management, compliance and associated regulatory documents. The model maps sentence pairs to a 768 dimensional dense vector space and can be used for tasks like clustering or semantic search as part of retrieval-augmented generation pipelines.

Evaluation Results

The model was evaluated on a held-out sample from the STAR-QA dataset (see below) using sentence-transformers.InformationRetrievalEvaluator. Reported metrics include cosine similarity of retrieved documents w/r/t ground truth P/R @ 3 candidates, as well as MRR @ 10, MAP @ 10 and NDCG @ 100. This fine-tuned model was also benchmarked against its base model using the same methodology.

Metric	STAR-QA Score	ALL-MPNET-BASE-V2 Score
Precision @ 3	0.315	0.215
Recall @ 3	0.324	0.223
MRR @ 10	0.887	0.578
NDCG @ 10	0.44	0.303
MAP @ 100	0.316	0.209

Training Data

The model was fine-tuned on a corpus of audit, risk-management, compliance and associated regulatory documents sourced from the public internet. Documents were cleaned and chunked into 2-sentence blocks. Each block was then sent to a state-of-the-art LLM with the following prompt: "Write a question about {document_topic} for which this is the answer: {block}"

The resulting question and its associated ground-truth answer (collectively a "pair") constitute a single training example for the fine-tuning step. The final model was fine-tuned on ~18K such pairs.

Training

The model was fine-tuned with the parameters:

DataLoader:

torch.utils.data.dataloader.DataLoader of length 634 with parameters:

{'batch_size': 16, 'sampler': 'torch.utils.data.sampler.RandomSampler', 'batch_sampler': 'torch.utils.data.sampler.BatchSampler'}

Loss:

sentence_transformers.losses.MultipleNegativesRankingLoss.MultipleNegativesRankingLoss with parameters:

{'scale': 20.0, 'similarity_fct': 'cos_sim'}

Parameters of the fit()-Method:

{
    "epochs": 1,
    "evaluation_steps": 50,
    "evaluator": "sentence_transformers.evaluation.InformationRetrievalEvaluator.InformationRetrievalEvaluator",
    "max_grad_norm": 1,
    "optimizer_class": "<class 'torch.optim.adamw.AdamW'>",
    "optimizer_params": {
        "lr": 2e-05
    },
    "scheduler": "WarmupLinear",
    "steps_per_epoch": null,
    "warmup_steps": 10000,
    "weight_decay": 0.01
}

Full Model Architecture

SentenceTransformer(
  (0): Transformer({'max_seq_length': 384, 'do_lower_case': False}) with Transformer model: MPNetModel 
  (1): Pooling({'word_embedding_dimension': 768, 'pooling_mode_cls_token': False, 'pooling_mode_mean_tokens': True, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False})
  (2): Normalize()
)

Citing & Authors

@misc{Theron_2024, 
  title={Sentence Transformer for Audit Retrieval Question-Answering (STAR-QA)},
  url={https://huggingface.co/dptrsa/STAR-QA},
  author={Theron, Daniel},
  year={2024},
  month={Feb}
}