Dataset card
Browse files
README.md
ADDED
|
@@ -0,0 +1,29 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
pretty_name: crit-eval reference packages
|
| 4 |
+
tags:
|
| 5 |
+
- crit
|
| 6 |
+
- crit-eval
|
| 7 |
+
- evaluation
|
| 8 |
+
---
|
| 9 |
+
|
| 10 |
+
# crit-eval reference packages
|
| 11 |
+
|
| 12 |
+
Reference scores for [crit-eval](https://github.com/crit-labs/crit-eval): each package is a checkpoint's prepared
|
| 13 |
+
Eval data (`prepared.json`) and the reference's scores of its quality texts (`reference/`, and
|
| 14 |
+
`reference-as-quantized/` for the `as_quantized` quality mode), computed once so a run needs only the engine.
|
| 15 |
+
|
| 16 |
+
```text
|
| 17 |
+
pip install crit-eval
|
| 18 |
+
crit-eval fetch --model MODEL_DIR --data EVAL_DATA_DIR
|
| 19 |
+
```
|
| 20 |
+
|
| 21 |
+
`packages.json` lists every package with the identities it was made under (model, suite, protocol, reference version)
|
| 22 |
+
and the sha256 of each file, which `fetch` checks before it writes anything. Packages are named by the first 16 hex
|
| 23 |
+
digits of the model's and the suite's sha256.
|
| 24 |
+
|
| 25 |
+
| Package | Checkpoint |
|
| 26 |
+
| --- | --- |
|
| 27 |
+
| `a005aa0e0d07c24d-901dbbb231b062d7` | unsloth Qwen3.8-27B NVFP4 (with `as_quantized` scores) |
|
| 28 |
+
| `37bf285c743cba71-901dbbb231b062d7` | nvidia Qwen3.8-Flash-Next NVFP4 |
|
| 29 |
+
| `07d03ee7ff9706fc-901dbbb231b062d7` | canada-quant DeepSeek-V4-Flash W4A16-FP8 |
|