adlumal/auslaw-embed-v1.0

AusLaw Embedding Model v1.0

This is a sentence-transformers model: It maps sentences & paragraphs to a 384 dimensional dense vector space and can be used for tasks like clustering or semantic search.

This model is a fine-tune of BAAI/bge-small-en using the HCA case law in the Open Australian Legal Corpus by Umar Butler. The PDF/OCR cases were not used.

The cases were split into < 512 context chunks using the bge-small-en tokeniser and semchunk.

mistralai/Mixtral-8x7B-Instruct-v0.1 was used to generate a legal question for each context chunk.

129,137 context-question pairs were used for training.

14,348 context-question pairs were used for evaluation (see the table below for results).

Using a 10% subset of the val dataset the following hit-rate performance was reached and is compared to the base model and OpenAI's default ada embedding model.

Model	Avg. hit-rate
BAAI/bge-small-en	89%
OpenAI	92%
adlumal/auslaw-embed-v1.0	97%

Usage (Sentence-Transformers)

Using this model becomes easy when you have sentence-transformers installed:

pip install -U sentence-transformers

Then you can use the model like this:

from sentence_transformers import SentenceTransformer
sentences = ["This is an example sentence", "Each sentence is converted"]

model = SentenceTransformer('adlumal/auslaw-embed-v1.0')
embeddings = model.encode(sentences)
print(embeddings)

Evaluation Results

The model was evauluated on 10% of the available data. The automated eval results for the final step are presented below.

Eval	Score
cos_sim-Accuracy@1	0.730206301
cos_sim-Accuracy@3	0.859562308
cos_sim-Accuracy@5	0.892737664
cos_sim-Accuracy@10	0.928352384
cos_sim-Precision@1	0.730206301
cos_sim-Recall@1	0.730206301
cos_sim-Precision@3	0.286520769
cos_sim-Recall@3	0.859562308
cos_sim-Precision@5	0.178547533
cos_sim-Recall@5	0.892737664
cos_sim-Precision@10	0.092835238
cos_sim-Recall@10	0.928352384
cos_sim-MRR@10	0.801075782
cos_sim-NDCG@10	0.832189447
cos_sim-MAP@100	0.803593645
dot_score-Accuracy@1	0.730136604
dot_score-Accuracy@3	0.859562308
dot_score-Accuracy@5	0.892737664
dot_score-Accuracy@10	0.928352384
dot_score-Precision@1	0.730136604
dot_score-Recall@1	0.730136604
dot_score-Precision@3	0.286520769
dot_score-Recall@3	0.859562308
dot_score-Precision@5	0.178547533
dot_score-Recall@5	0.892737664
dot_score-Precision@10	0.092835238
dot_score-Recall@10	0.928352384
dot_score-MRR@10	0.801040934
dot_score-NDCG@10	0.832163724
dot_score-MAP@100	0.803558796

Training

The model was trained with the parameters:

DataLoader:

torch.utils.data.dataloader.DataLoader of length 2583 with parameters:

{'batch_size': 50, 'sampler': 'torch.utils.data.sampler.SequentialSampler', '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": 2,
    "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": 516,
    "weight_decay": 0.01
}

Full Model Architecture

SentenceTransformer(
  (0): Transformer({'max_seq_length': 512, 'do_lower_case': True}) with Transformer model: BertModel 
  (1): Pooling({'word_embedding_dimension': 384, 'pooling_mode_cls_token': True, 'pooling_mode_mean_tokens': False, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False})
  (2): Normalize()
)

Citing & Authors

@misc{malec-2024-auslaw-embed-v1,
    author = {Malec, Adrian Lucas},
    year = {2024},
    title = {AusLaw Embedding v1.0},
    publisher = {Hugging Face},
    version = {1.0},
    url = {https://huggingface.co/adlumal/auslaw-embed-v1.0}
}