stonarini commited on
Commit
1ccfff1
·
verified ·
1 Parent(s): 760d207

feat: grepac-e5-small v1

Browse files
.gitattributes CHANGED
@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ tokenizer.json filter=lfs diff=lfs merge=lfs -text
1_Pooling/config.json ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ {
2
+ "embedding_dimension": 384,
3
+ "pooling_mode": "mean",
4
+ "include_prompt": true
5
+ }
2_Normalize/config.json ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ {
2
+ "module_input_name": "sentence_embedding",
3
+ "module_output_name": "sentence_embedding"
4
+ }
README.md ADDED
@@ -0,0 +1,89 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language: [en]
3
+ license: mit
4
+ library_name: sentence-transformers
5
+ pipeline_tag: sentence-similarity
6
+ base_model: intfloat/multilingual-e5-small
7
+ tags: [sentence-transformers, retrieval, legal, agriculture, eu-law, cap]
8
+ ---
9
+
10
+ # grepac-e5-small
11
+
12
+ `intfloat/multilingual-e5-small` fine-tuned to find the right article of the EU Common
13
+ Agricultural Policy regulations for a farmer's or advisor's question. It powers
14
+ [grePAC](https://huggingface.co/spaces/stonarini/grepac). Code and training pipeline: [github.com/stonarini/grepac](https://github.com/stonarini/grepac).
15
+
16
+ Corpus: Regulation (EU) 2021/2115 (consolidated 2026-08-18) and Regulation (EU) 2021/2116
17
+ (consolidated 2024-05-25), articles only, split into 911 paragraphs.
18
+
19
+ ## Usage
20
+
21
+ Same prefixes as the base model.
22
+
23
+ ```python
24
+ from sentence_transformers import SentenceTransformer
25
+ model = SentenceTransformer("stonarini/grepac-e5-small")
26
+ q = model.encode(["query: When will I receive my CAP payments?"], normalize_embeddings=True)
27
+ p = model.encode(["passage: Article 44 Payment to beneficiaries\n..."], normalize_embeddings=True)
28
+ print(q @ p.T)
29
+ ```
30
+
31
+ Passages were embedded as `passage: Article <n> — <title>\n<lead-in>\n<paragraph text>`.
32
+
33
+ ## Training data
34
+
35
+ Generated automatically, no hand-written training examples:
36
+
37
+ - Mistral 7B Instruct v0.3 (Q4_K_M, local via Ollama) wrote 4 questions per paragraph
38
+ (2 farmer-style, 1 keyword search, 1 advisor-style), plus rule-based queries from article
39
+ titles and from the definitions.
40
+ - Filters: a question was dropped if the base model did not rank its article in the top 20
41
+ (likely a misreading, 257 dropped), if it was too similar to a test question (cosine ≥ 0.93,
42
+ 153 dropped), or if it was a duplicate (10).
43
+ - Each question got a hard negative: a paragraph from another article that the base model
44
+ scored close to the right one.
45
+ - Result: 2,933 triplets over 239 articles; 287 questions from 26 other articles held out for
46
+ validation.
47
+
48
+ Training: MultipleNegativesRankingLoss with hard negatives, no duplicate texts in a batch,
49
+ 2 epochs, batch 32, learning rate 2e-5, max length 256, on a 4-core laptop CPU (73 minutes).
50
+
51
+ ## Evaluation
52
+
53
+ 60 hand-written and hand-checked questions (38 practical, 18 technical, 4 in Italian, French,
54
+ German and Spanish), each with the article(s) that answer it. None were produced by the
55
+ training LLM. Scored at article level.
56
+
57
+ | | right article first | in top 3 | in top 5 | MRR@10 |
58
+ |---|---|---|---|---|
59
+ | base `multilingual-e5-small` | 0.667 | 0.833 | 0.867 | 0.750 |
60
+ | **grepac-e5-small** | **0.717** | **0.900** | **0.900** | **0.797** |
61
+
62
+ By question type, right article in the top 5: practical 0.816 → 0.868, technical 1.000 → 1.000,
63
+ other languages 0.750 → 0.750.
64
+
65
+ On the 287 held-out generated questions: right article first 0.735 → 0.794, top 5 0.937 → 0.962.
66
+
67
+ **How much to trust this:** 60 questions is a small test. Question by question, 11 got better
68
+ and 8 got worse, which is not statistically significant (sign test p ≈ 0.65). The gains are
69
+ concentrated on farmer-style wording (for example "What counts as a farmer?" moved from rank 31
70
+ to 2), the losses are mostly small. Treat this version as a working first release; it will be
71
+ re-evaluated on real searches from the site.
72
+
73
+ Things that did not help, measured on the same 60 questions: adding a fixed "EU Common
74
+ Agricultural Policy" context or synonyms to the query, and adding the chapter headings or the
75
+ regulation name to the passages.
76
+
77
+ ## Known limitations
78
+
79
+ - The LLM misread some passages. Example: for the capping of payments (2115 Art. 17) it wrote
80
+ questions about farm *income* above EUR 60,000, so questions about the payment cap for large
81
+ farms got worse.
82
+ - Trained on English questions only. The other languages were checked on 4 questions.
83
+ - Annexes are not in the corpus (for example the GAEC standards in Annex III of 2021/2115).
84
+ - Built for this corpus; not tested on other legal texts.
85
+
86
+ ## License
87
+
88
+ MIT, as the base model. Corpus: EUR-Lex, © European Union, reused under the Commission's
89
+ reuse policy. Training questions generated with Mistral 7B Instruct v0.3 (Apache-2.0).
config.json ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_cross_attention": false,
3
+ "architectures": [
4
+ "BertModel"
5
+ ],
6
+ "attention_probs_dropout_prob": 0.1,
7
+ "bos_token_id": 0,
8
+ "classifier_dropout": null,
9
+ "dtype": "float32",
10
+ "eos_token_id": 2,
11
+ "hidden_act": "gelu",
12
+ "hidden_dropout_prob": 0.1,
13
+ "hidden_size": 384,
14
+ "initializer_range": 0.02,
15
+ "intermediate_size": 1536,
16
+ "is_decoder": false,
17
+ "layer_norm_eps": 1e-12,
18
+ "max_position_embeddings": 512,
19
+ "model_type": "bert",
20
+ "num_attention_heads": 12,
21
+ "num_hidden_layers": 12,
22
+ "pad_token_id": 1,
23
+ "position_embedding_type": "absolute",
24
+ "tie_word_embeddings": true,
25
+ "tokenizer_class": "XLMRobertaTokenizer",
26
+ "transformers_version": "5.22.1",
27
+ "type_vocab_size": 2,
28
+ "use_cache": false,
29
+ "vocab_size": 250037
30
+ }
config_sentence_transformers.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "__version__": {
3
+ "pytorch": "2.15.0+cu132",
4
+ "sentence_transformers": "6.2.0",
5
+ "transformers": "5.22.1"
6
+ },
7
+ "default_prompt_name": null,
8
+ "model_type": "SentenceTransformer",
9
+ "prompts": {
10
+ "document": "",
11
+ "query": ""
12
+ },
13
+ "similarity_fn_name": "cosine"
14
+ }
model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:81485d735cbdb202703ed7b2847ea0474ddf92f62c8748fb88915c12f925708a
3
+ size 470637416
modules.json ADDED
@@ -0,0 +1,20 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "idx": 0,
4
+ "name": "0",
5
+ "path": "",
6
+ "type": "sentence_transformers.base.modules.transformer.Transformer"
7
+ },
8
+ {
9
+ "idx": 1,
10
+ "name": "1",
11
+ "path": "1_Pooling",
12
+ "type": "sentence_transformers.sentence_transformer.modules.pooling.Pooling"
13
+ },
14
+ {
15
+ "idx": 2,
16
+ "name": "2",
17
+ "path": "2_Normalize",
18
+ "type": "sentence_transformers.base.modules.normalize.Normalize"
19
+ }
20
+ ]
sentence_bert_config.json ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "transformer_task": "feature-extraction",
3
+ "modality_config": {
4
+ "text": {
5
+ "method": "forward",
6
+ "method_output_name": "last_hidden_state"
7
+ }
8
+ },
9
+ "module_output_name": "token_embeddings"
10
+ }
tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fb02bfe42ffe0fe115de5c74f76c6480d730e60cc1f6ae567ff4d82cb305a113
3
+ size 17082988
tokenizer_config.json ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_prefix_space": true,
3
+ "backend": "tokenizers",
4
+ "bos_token": "<s>",
5
+ "clean_up_tokenization_spaces": true,
6
+ "cls_token": "<s>",
7
+ "eos_token": "</s>",
8
+ "is_local": false,
9
+ "local_files_only": true,
10
+ "mask_token": "<mask>",
11
+ "model_max_length": 256,
12
+ "pad_token": "<pad>",
13
+ "sep_token": "</s>",
14
+ "sp_model_kwargs": {},
15
+ "tokenizer_class": "XLMRobertaTokenizer",
16
+ "unk_token": "<unk>"
17
+ }