Upload README.md with huggingface_hub
Browse files
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
|
| 102 |
-
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 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 |
-
|
|
|
|
|
|
|
| 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 |
+

|
| 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"])
|