Update README.md
Browse files
README.md
CHANGED
@@ -16,7 +16,7 @@ language:
|
|
16 |
|
17 |
# Weaver Distilled for MMLU-Pro
|
18 |
|
19 |
-
|
20 |
|
21 |
## Model Details
|
22 |
|
@@ -26,14 +26,6 @@ A distilled cross-encoder model that captures 98.7% of Weaver's accuracy while r
|
|
26 |
- **Training Data**: MMLU-Pro problems with Weaver scores from 35 LM judges and reward models
|
27 |
- **Task**: Binary classification for answer correctness prediction
|
28 |
|
29 |
-
## Performance
|
30 |
-
|
31 |
-
On MMLU-Pro with Llama 3.1 70B generations:
|
32 |
-
<!-- TODO: Update with actual performance numbers -->
|
33 |
-
- **Weaver (Full)**: XX.X% accuracy, high compute cost
|
34 |
-
- **Weaver (Distilled)**: XX.X% accuracy, 99.97% compute reduction
|
35 |
-
- **Majority Voting**: XX.X% accuracy
|
36 |
-
|
37 |
## Quick Start
|
38 |
|
39 |
```python
|
|
|
16 |
|
17 |
# Weaver Distilled for MMLU-Pro
|
18 |
|
19 |
+
This is a distilled cross-encoder model based on ModernBERT-large, trained to predict the correctness of answers on MMLU Pro. This specialized verifier was trained on Weaver scores aggregated over 35 different verifiers and reward models.
|
20 |
|
21 |
## Model Details
|
22 |
|
|
|
26 |
- **Training Data**: MMLU-Pro problems with Weaver scores from 35 LM judges and reward models
|
27 |
- **Task**: Binary classification for answer correctness prediction
|
28 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
29 |
## Quick Start
|
30 |
|
31 |
```python
|