Ill-Ness commited on
Commit
830cdf4
·
verified ·
1 Parent(s): 0e11747

Upload Benchmark

Browse files
Files changed (3) hide show
  1. README.md +64 -30
  2. dataset_info.json +3 -2
  3. eval.yaml +1 -1
README.md CHANGED
@@ -15,7 +15,8 @@ pretty_name: ParseEmbed
15
  size_categories:
16
  - n<1K
17
  tags:
18
- - mteb
 
19
  - text
20
  - retrieval
21
  - hard-negatives
@@ -116,27 +117,11 @@ dataset_info:
116
 
117
  # ParseEmbed
118
 
119
- ParseEmbed is a compact retrieval benchmark for testing whether embedding
120
- models actually preserve parse-sensitive meaning, not just broad topic
121
- similarity.
122
 
123
- Many embedding benchmarks reward models for finding text that is generally
124
- about the same subject. ParseEmbed is designed to be harder: every query has
125
- one correct document and several hard negatives that reuse almost the same
126
- words, entities, numbers, and structure. A model only scores well if it notices
127
- the detail that changes the answer.
128
-
129
- The benchmark focuses on three common failure modes:
130
-
131
- - `mean`: semantic details such as negation, numeric values, scope, conditions,
132
- and exceptions.
133
- - `text_formatting`: meaning carried by Markdown-like formatting, including
134
- code spans, headings, quotes, emphasis, and strike-through.
135
- - `table`: row and column grounding in compact Markdown tables.
136
-
137
- The purpose of ParseEmbed is diagnostic. It helps model builders see whether an
138
- embedding model is robust enough for retrieval over structured or semi-structured
139
- documents, where a tiny parsing mistake can retrieve the wrong answer.
140
 
141
  ## Tasks
142
 
@@ -153,6 +138,47 @@ documents, where a tiny parsing mistake can retrieve the wrong answer.
153
  - `corpus.jsonl`: retrieval corpus.
154
  - `queries.jsonl`: retrieval queries.
155
  - `qrels.jsonl`: relevance judgments.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
156
 
157
  ## Eval Results
158
 
@@ -169,6 +195,19 @@ Model repositories can report results with `.eval_results/parseembed.yaml`:
169
 
170
  Use task IDs `mean`, `text_formatting`, and `table`.
171
 
 
 
 
 
 
 
 
 
 
 
 
 
 
172
  ## Usage
173
 
174
  ```python
@@ -190,16 +229,11 @@ most of the vocabulary.
190
  No private or external source text is used. The benchmark is synthetic and is
191
  intended as a diagnostic stress test alongside natural retrieval benchmarks.
192
 
193
- ## Interpreting Scores
194
-
195
- High scores mean the model can separate near-identical documents using the
196
- specific detail requested by the query. Low scores usually mean the model is
197
- overweighting topical similarity or lexical overlap.
198
 
199
- ParseEmbed should not be treated as a replacement for broad retrieval
200
- benchmarks. It is best used as a focused stress test for document parsing,
201
- enterprise search, RAG systems, table retrieval, and formatting-sensitive text
202
- retrieval.
203
 
204
  ## Size
205
 
 
15
  size_categories:
16
  - n<1K
17
  tags:
18
+ - benchmark
19
+ - evaluation
20
  - text
21
  - retrieval
22
  - hard-negatives
 
117
 
118
  # ParseEmbed
119
 
120
+ Hard, parse-sensitive retrieval evaluation for embedding models.
 
 
121
 
122
+ ParseEmbed is a compact benchmark for embedding models. It tests whether a
123
+ model can retrieve the exact correct document when hard negatives share nearly
124
+ all surface tokens with the answer.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
125
 
126
  ## Tasks
127
 
 
138
  - `corpus.jsonl`: retrieval corpus.
139
  - `queries.jsonl`: retrieval queries.
140
  - `qrels.jsonl`: relevance judgments.
141
+ - `evaluate_parseembed.py`: reference evaluator for embedding models.
142
+ - `requirements.txt`: evaluator dependencies.
143
+
144
+ ## Quick Start
145
+
146
+ Install the evaluator dependencies:
147
+
148
+ ```bash
149
+ pip install -r requirements.txt
150
+ ```
151
+
152
+ Run a local evaluation:
153
+
154
+ ```bash
155
+ python evaluate_parseembed.py --model sentence-transformers/all-MiniLM-L6-v2 --task all --k 10 --output results.json
156
+ ```
157
+
158
+ The evaluator reports `ndcg_at_10`, `recall_at_10`, and `mrr_at_10` for each
159
+ task and an average across tasks.
160
+
161
+ ## Evaluation Protocol
162
+
163
+ ParseEmbed is evaluated as dense retrieval:
164
+
165
+ 1. Embed every document in `corpus.jsonl`.
166
+ 2. Embed the queries from one benchmark split.
167
+ 3. Rank corpus documents by cosine similarity.
168
+ 4. Score the rank of each query's `positive_doc_id`.
169
+
170
+ The official leaderboard metric should be `ndcg_at_10`. `recall_at_10` and
171
+ `mrr_at_10` are included as secondary diagnostics.
172
+
173
+ ## Data Format
174
+
175
+ Each benchmark split contains one JSON object per query with these fields:
176
+ `id`, `query`, `positive_doc_id`, `positive_text`, `hard_negative_doc_ids`,
177
+ `hard_negative_texts`, `answer`, `style`, and `difficulty`.
178
+
179
+ The corpus is stored separately in `corpus.jsonl`, and `qrels.jsonl` provides
180
+ standard retrieval relevance judgments with `query-id`, `corpus-id`, and
181
+ `score`.
182
 
183
  ## Eval Results
184
 
 
195
 
196
  Use task IDs `mean`, `text_formatting`, and `table`.
197
 
198
+ ## Benchmark Registration
199
+
200
+ This repository is prepared for Hugging Face's beta Benchmark system:
201
+
202
+ - It contains a root `eval.yaml`.
203
+ - `evaluation_framework` is set to `parseembed`.
204
+ - The benchmark has three task leaderboards: `mean`, `text_formatting`, and
205
+ `table`.
206
+ - A working reference evaluator is included in `evaluate_parseembed.py`.
207
+
208
+ Because Hugging Face's Benchmark feature is allow-listed, `parseembed` must be
209
+ added by the Hugging Face team before the official Benchmark tag appears.
210
+
211
  ## Usage
212
 
213
  ```python
 
229
  No private or external source text is used. The benchmark is synthetic and is
230
  intended as a diagnostic stress test alongside natural retrieval benchmarks.
231
 
232
+ ## Reproducibility
 
 
 
 
233
 
234
+ ParseEmbed is fully deterministic. Regenerating from `generate_parseembed.py`
235
+ with the same seed produces the same queries, corpus IDs, hard negatives, and
236
+ relevance judgments.
 
237
 
238
  ## Size
239
 
dataset_info.json CHANGED
@@ -32,7 +32,8 @@
32
  "n<1K"
33
  ],
34
  "tags": [
35
- "mteb",
 
36
  "text",
37
  "retrieval",
38
  "hard-negatives",
@@ -69,4 +70,4 @@
69
  ],
70
  "corpus_documents": 2880
71
  }
72
- }
 
32
  "n<1K"
33
  ],
34
  "tags": [
35
+ "benchmark",
36
+ "evaluation",
37
  "text",
38
  "retrieval",
39
  "hard-negatives",
 
70
  ],
71
  "corpus_documents": 2880
72
  }
73
+ }
eval.yaml CHANGED
@@ -3,7 +3,7 @@ description: >
3
  ParseEmbed is a hard retrieval benchmark for embedding models. It measures
4
  semantic scope, formatting-sensitive meaning, and table cell grounding under
5
  near-duplicate hard negatives.
6
- evaluation_framework: mteb
7
 
8
  tasks:
9
  - id: mean
 
3
  ParseEmbed is a hard retrieval benchmark for embedding models. It measures
4
  semantic scope, formatting-sensitive meaning, and table cell grounding under
5
  near-duplicate hard negatives.
6
+ evaluation_framework: parseembed
7
 
8
  tasks:
9
  - id: mean