File size: 2,025 Bytes
1ccfff1
 
 
 
 
 
 
 
 
 
 
8940247
 
 
 
1ccfff1
 
 
8940247
1ccfff1
 
 
8940247
1ccfff1
 
 
 
 
 
8940247
1ccfff1
8940247
1ccfff1
8940247
 
1ccfff1
8940247
1ccfff1
 
 
8940247
1ccfff1
 
 
8940247
 
1ccfff1
8940247
1ccfff1
 
 
 
8940247
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
---
language: [en]
license: mit
library_name: sentence-transformers
pipeline_tag: sentence-similarity
base_model: intfloat/multilingual-e5-small
tags: [sentence-transformers, retrieval, legal, agriculture, eu-law, cap]
---

# grepac-e5-small

[`intfloat/multilingual-e5-small`](https://huggingface.co/intfloat/multilingual-e5-small)
fine-tuned on CPU to find the right article of the EU Common Agricultural Policy (CAP)
regulations for a question. It powers [grePAC](https://huggingface.co/spaces/stonarini/grepac).
Code and training pipeline: [github.com/stonarini/grepac](https://github.com/stonarini/grepac).

## Usage

Same `query: ` / `passage: ` prefixes as the base model.

```python
from sentence_transformers import SentenceTransformer

model = SentenceTransformer("stonarini/grepac-e5-small")
q = model.encode(["query: When will I receive my CAP payments?"], normalize_embeddings=True)
p = model.encode(["passage: Article 44 Payment to beneficiaries\n..."], normalize_embeddings=True)
print(q @ p.T)
```

## Training

The training data is generated, with no hand-written examples:

- a local LLM (Mistral 7B Instruct v0.3) writes questions for each paragraph, plus rule-based queries from titles and definitions
- each question gets a hard negative, a paragraph from another article that looks similar

~3000 training examples over +200 articles. MultipleNegativesRankingLoss, 2 epochs, batch 32, learning rate 2e-5, 73 minutes on a 4-core laptop CPU.

## Evaluation

60 hand-written and hand-checked questions, each with the articles that answer it. Scored per article.

| | right article first | in top 3 | in top 5 | MRR@10 |
|---|---|---|---|---|
| base model | 0.667 | 0.833 | 0.867 | 0.750 |
| grepac-e5-small | 0.717 | 0.900 | 0.900 | 0.797 |

The test is too small to be conclusive: 11 questions improved and 8 got worse. Real searches from the site will be the next test set.


## License

MIT, like the base model. Regulation texts: EUR-Lex, © European Union, reused under the Commission's reuse policy.