translorentz commited on
Commit
1a3c236
·
verified ·
1 Parent(s): ef0a75e

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +29 -8
README.md CHANGED
@@ -23,10 +23,14 @@ configs:
23
  data_files:
24
  - split: test
25
  path: en/test.jsonl
 
 
26
  - config_name: zh
27
  data_files:
28
  - split: test
29
  path: zh/test.jsonl
 
 
30
  ---
31
 
32
  # OPTIC-Bench
@@ -98,13 +102,20 @@ to the questions this benchmark was built to settle. The full report is in
98
  adversarial text was imaged, although an older model tested during
99
  development was susceptible under the combined-image condition.
100
 
101
- These are pilot figures from very small per-cell samples, so the percentages
102
- are directional and should be reproduced at the full per-cell count of the
103
- dataset before being quoted as rankings. The qualitative conclusions, that
104
- exact retrieval collapses from a dense image whilst reasoning is preserved,
105
- hold with statistical significance even at this scale. See `RESULTS.md` for
106
- the significance tests, the character-level versus field-level detail, and the
107
- full caveats.
 
 
 
 
 
 
 
108
 
109
  ## Fields
110
 
@@ -128,11 +139,21 @@ the Hub.
128
 
129
  ## Usage
130
 
 
 
 
 
 
 
 
 
131
  ```python
132
  import json
133
  from datasets import load_dataset
134
 
135
- ds = load_dataset("translorentz/vision-token-compression-bench", "en", split="test") # or "zh"
 
 
136
  row = ds[0]
137
  expected = json.loads(row["expected"])
138
  meta = json.loads(row["meta"])
 
23
  data_files:
24
  - split: test
25
  path: en/test.jsonl
26
+ - split: core
27
+ path: en/core.jsonl
28
  - config_name: zh
29
  data_files:
30
  - split: test
31
  path: zh/test.jsonl
32
+ - split: core
33
+ path: zh/core.jsonl
34
  ---
35
 
36
  # OPTIC-Bench
 
102
  adversarial text was imaged, although an older model tested during
103
  development was susceptible under the combined-image condition.
104
 
105
+ These conclusions have been reproduced on the 322-instance core subset for
106
+ all three models, where the retrieval collapse is significant at an exact
107
+ McNemar probability below ten to the power minus twelve. The figure below
108
+ places each task family on the trade-off between the tokens saved and the
109
+ accuracy lost on long dense documents. Reasoning and instruction keep their
110
+ accuracy whilst saving tokens, whereas exact retrieval and transcription lose
111
+ almost everything.
112
+
113
+ ![The economy-fidelity trade-off](charts/4_economy_vs_fidelity.png)
114
+
115
+ The token-economy percentages come from small long-document samples and are
116
+ directional. See `RESULTS.md` for the per-model core tables, the four
117
+ performance charts, the significance tests, the character-level versus
118
+ field-level detail, and the full caveats.
119
 
120
  ## Fields
121
 
 
139
 
140
  ## Usage
141
 
142
+ Each language configuration provides two splits. The `test` split is the full
143
+ set. The `core` split is a deterministic 322-instance subset weighted towards
144
+ the harder tiers, recommended as the default for routine evaluation because a
145
+ full four-condition sweep of it is about 1,150 calls per model rather than
146
+ 6,100. The core is a strict subset of `test`, so its instances, ground truth
147
+ and images are identical to the full set. Use `test` when tighter confidence
148
+ intervals are wanted.
149
+
150
  ```python
151
  import json
152
  from datasets import load_dataset
153
 
154
+ # recommended default: the smaller core split
155
+ ds = load_dataset("translorentz/vision-token-compression-bench", "en", split="core")
156
+ # or the full set: split="test"
157
  row = ds[0]
158
  expected = json.loads(row["expected"])
159
  meta = json.loads(row["meta"])