datalyes commited on
Commit
a59a74c
·
verified ·
1 Parent(s): 780f401

Added MTEB availability and partial reproduction steps

Browse files
Files changed (1) hide show
  1. README.md +121 -35
README.md CHANGED
@@ -25,59 +25,145 @@ configs:
25
 
26
  # **DAPFAM** dataset
27
 
28
- For more details on the dataset construction and full set of experimentations and results, see the accompanying paper: **Ayaou et al., 2025 “DAPFAM: A Domain-Aware Family-level Dataset to benchmark cross domain patent retrieval”[(Here)](https://doi.org/10.48550/arXiv.2506.22141) .**
29
 
 
30
 
31
- ## Summary
 
 
 
 
32
 
33
- DAPFAM provides **1 247 domain balanced full-text query patent families** and **45 336 full-text target families** with forward/backward‑citation relevance labels (≈ 50 K pairs). Each relevant link is explicitly marked **in‑domain** or **out‑of‑domain** according to IPC 3‑char overlap, enabling rigorous cross‑domain evaluation.
34
 
35
- * Full text **(title · abstract · claims · description)** plus rich metadata for *every* family.
36
- * Multi‑jurisdictional, English‑only text (families may originate in US, JP, EP, CN, …).
37
- * Parquet qrel file: `qrels_all.parquet`.
38
 
39
- ## Dataset Structure
40
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
41
  ```
42
- corpus.parquet # 45 336 rows, targets – every original column from the paper
43
- queries.parquet # 1 247 rows, queries – same columns + abstract_keywords
44
- qrels_all.parquet # (all | in | out) four‑column tables → query_id · relevant_id · relevance_score · domain_rel
45
- ```
46
 
47
- ## How to load
 
 
 
 
48
 
49
  ```python
50
  from datasets import load_dataset
51
 
52
- #According to your usage, you might not need to load all 3 subsets
 
 
 
53
 
54
- dc = load_dataset("datalyes/DAPFAM_patent", "corpus")
 
 
 
 
55
 
56
- dq = load_dataset("datalyes/DAPFAM_patent", "queries")
57
 
58
- dr = load_dataset("datalyes/DAPFAM_patent", "relations")
59
- ```
 
 
60
 
61
- ## Citation
62
 
63
- If you find our paper or dataset helpful, please consider citing as follows:
64
 
 
65
  ```
66
  @misc{ayaou2025dapfamdomainawarefamilyleveldataset,
67
- title={DAPFAM: A Domain-Aware Family-level Dataset to benchmark cross domain patent retrieval},
68
- author={Iliass Ayaou and Denis Cavallucci and Hicham Chibane},
69
- year={2025},
70
- eprint={2506.22141},
71
- archivePrefix={arXiv},
72
- primaryClass={cs.CL},
73
- url={https://arxiv.org/abs/2506.22141},
74
  }
75
- ```
76
-
77
- ## Quick Stats
78
-
79
- * **Queries**: 1,247
80
- * **Corpus (targets)**: 45,336
81
- * **Qrels (all)**: 49,869
82
- * **Qrels (in)**: 19,736
83
- * **Qrels (out)**: 5,193
 
25
 
26
  # **DAPFAM** dataset
27
 
28
+ > **What’s new (Sept 2025)** **DAPFAM patent family retrieval tasks are now in MTEB.** 18 tasks (ALL / IN / OUT × 3 query views × 3 target views) are available, including the 6 main ones used in our paper. You can benchmark any model with a single script and reproduce the paper’s results by selecting the same encoder (Snowflake/snowflake-arctic-embed-m-v2.0). Our paper used int8 quantization for hardware reasons; results may vary very slightly (not significantly) if you run in float16/32.
29
 
30
+ ### DAPFAM — A Domain‑Aware Family‑level Dataset to benchmark cross‑domain patent retrieval
31
 
32
+ **License:** CC‑BY‑NC‑SA‑4.0
33
+ **Tasks:** text‑retrieval (patent family prior‑art retrieval)
34
+ **Languages:** English (eng‑Latn)
35
+ **Evaluation date span:** 1964‑06‑26 → 2023‑06‑20
36
+ **Cite:** Ayaou et al., 2025 — _DAPFAM: A Domain‑Aware Family‑level Dataset to benchmark cross‑domain patent retrieval_ (arXiv:2506.22141)
37
 
38
+ ---
39
 
40
+ ### Summary
 
 
41
 
42
+ **DAPFAM** provides **1,247 query patent families** and **45,336 target families** with **citation‑based relevance** and explicit **domain labels** (IN/OUT). Each positive pair is IN‑domain if query and target share at least one IPC3 code, OUT‑domain otherwise. Text is at **family‑level full text** (title, abstract, claims, description). Supports both **document-** and **passage‑level** retrieval.
43
 
44
+ **What makes DAPFAM different?**
45
+ - **Explicit domain partitions** (IN vs OUT) → enables true cross‑domain evaluation.
46
+ - **Family‑level aggregation** → reduces cross‑jurisdiction redundancy.
47
+ - **Compute‑aware** → Small enough to support passage level experimentations on consumer-grade hardware.
48
+
49
+ ---
50
+
51
+ ### Benchmark DAPFAM on MTEB
52
+
53
+ **18 retrieval tasks** have been added (ALL / IN / OUT × 3 query × 3 target field views). Six of them were directly evaluated in the paper.
54
+
55
+ #### Task naming scheme
56
+ - Query view: **TA** (Title+Abstract) or **TAC** (Title+Abstract+Claims)
57
+ - Target view: **TA**, **TAC**, or **FullText** (adds description)
58
+ - Subsets: **ALL**, **IN**, **OUT** (IPC overlap filtering)
59
+
60
+ #### Task list (18 total)
61
+
62
+ **ALL**
63
+ - `DAPFAMAllTitlAbsToTitlAbsRetrieval`
64
+ - `DAPFAMAllTitlAbsToTitlAbsClmRetrieval` **(in-paper)**
65
+ - `DAPFAMAllTitlAbsToFullTextRetrieval`
66
+ - `DAPFAMAllTitlAbsClmToTitlAbsRetrieval`
67
+ - `DAPFAMAllTitlAbsClmToTitlAbsClmRetrieval` **(in-paper)**
68
+ - `DAPFAMAllTitlAbsClmToFullTextRetrieval`
69
+
70
+ **IN**
71
+ - `DAPFAMInTitlAbsToTitlAbsRetrieval`
72
+ - `DAPFAMInTitlAbsToTitlAbsClmRetrieval` **(in-paper)**
73
+ - `DAPFAMInTitlAbsToFullTextRetrieval`
74
+ - `DAPFAMInTitlAbsClmToTitlAbsRetrieval`
75
+ - `DAPFAMInTitlAbsClmToTitlAbsClmRetrieval` **(in-paper)**
76
+ - `DAPFAMInTitlAbsClmToFullTextRetrieval`
77
+
78
+ **OUT**
79
+ - `DAPFAMOutTitlAbsToTitlAbsRetrieval`
80
+ - `DAPFAMOutTitlAbsToTitlAbsClmRetrieval` **(in-paper)**
81
+ - `DAPFAMOutTitlAbsToFullTextRetrieval`
82
+ - `DAPFAMOutTitlAbsClmToTitlAbsRetrieval`
83
+ - `DAPFAMOutTitlAbsClmToTitlAbsClmRetrieval` **(in-paper)**
84
+ - `DAPFAMOutTitlAbsClmToFullTextRetrieval`
85
+
86
+ #### Quick start — run all tasks
87
+ ```python
88
+ import mteb
89
+ from sentence_transformers import SentenceTransformer
90
+
91
+ model_name = "Snowflake/snowflake-arctic-embed-m-v2.0"
92
+ model = SentenceTransformer(model_name, trust_remote_code=True,
93
+ model_kwargs={"torch_dtype":"float16"}).cuda().eval()
94
+
95
+ task_names = [
96
+ # ALL
97
+ "DAPFAMAllTitlAbsToTitlAbsRetrieval",
98
+ "DAPFAMAllTitlAbsToTitlAbsClmRetrieval",
99
+ "DAPFAMAllTitlAbsToFullTextRetrieval",
100
+ "DAPFAMAllTitlAbsClmToTitlAbsRetrieval",
101
+ "DAPFAMAllTitlAbsClmToTitlAbsClmRetrieval",
102
+ "DAPFAMAllTitlAbsClmToFullTextRetrieval",
103
+ # IN
104
+ "DAPFAMInTitlAbsToTitlAbsRetrieval",
105
+ "DAPFAMInTitlAbsToTitlAbsClmRetrieval",
106
+ "DAPFAMInTitlAbsToFullTextRetrieval",
107
+ "DAPFAMInTitlAbsClmToTitlAbsRetrieval",
108
+ "DAPFAMInTitlAbsClmToTitlAbsClmRetrieval",
109
+ "DAPFAMInTitlAbsClmToFullTextRetrieval",
110
+ # OUT
111
+ "DAPFAMOutTitlAbsToTitlAbsRetrieval",
112
+ "DAPFAMOutTitlAbsToTitlAbsClmRetrieval",
113
+ "DAPFAMOutTitlAbsToFullTextRetrieval",
114
+ "DAPFAMOutTitlAbsClmToTitlAbsRetrieval",
115
+ "DAPFAMOutTitlAbsClmToTitlAbsClmRetrieval",
116
+ "DAPFAMOutTitlAbsClmToFullTextRetrieval",
117
+ ]
118
+
119
+ tasks = mteb.get_tasks(tasks=task_names)
120
+ results = mteb.MTEB(tasks=tasks).run(
121
+ model,
122
+ output_folder=f"mteb_res/{model_name}",
123
+ encode_kwargs={"batch_size": 16, "prompt_name": None}
124
+ )
125
+ print(results)
126
  ```
 
 
 
 
127
 
128
+ > To reproduce the **paper’s reported MTEB-compatible results**, restrict to the six **in-paper tasks** listed above. The encoder was run with int8 quantization in the paper; float16 runs on GPU may differ slightly.
129
+
130
+ ---
131
+
132
+ ### How to Load the Dataset
133
 
134
  ```python
135
  from datasets import load_dataset
136
 
137
+ dc = load_dataset("datalyes/DAPFAM_patent", "corpus") # 45,336 targets
138
+ dq = load_dataset("datalyes/DAPFAM_patent", "queries") # 1,247 queries
139
+ dr = load_dataset("datalyes/DAPFAM_patent", "relations") # qrels: all/in/out
140
+ ```
141
 
142
+ **Counts**
143
+ - Queries: **1,247**
144
+ - Targets: **45,336**
145
+ - Qrels (all): **≈49,869** (positives + sampled negatives)
146
+ - Positive qrels: **IN ~19,736**, **OUT ~5,193**
147
 
148
+ ---
149
 
150
+ ### Evaluation choices
151
+ - Metrics: **NDCG@100** (primary), **Recall@100** (secondary).
152
+ - Document-level views in MTEB; paper also explores **passage-level** retrieval and **RRF fusion** separately.
153
+ - Encoder: `Snowflake/snowflake-arctic-embed-m-v2.0`; in-paper runs quantized to int8 for efficiency.
154
 
155
+ ---
156
 
 
157
 
158
+ ### Citation
159
  ```
160
  @misc{ayaou2025dapfamdomainawarefamilyleveldataset,
161
+ title={DAPFAM: A Domain-Aware Family-level Dataset to benchmark cross domain patent retrieval},
162
+ author={Iliass Ayaou and Denis Cavallucci and Hicham Chibane},
163
+ year={2025},
164
+ eprint={2506.22141},
165
+ archivePrefix={arXiv},
166
+ primaryClass={cs.CL},
167
+ url={https://arxiv.org/abs/2506.22141},
168
  }
169
+ ```