Datasets:
card: names the model and v1, non-affiliation line
Browse files- README.md +9 -8
- SHA256SUMS +1 -2
README.md
CHANGED
|
@@ -44,7 +44,8 @@ configs:
|
|
| 44 |
|
| 45 |
> PRIVATE DRAFT (not released). Built locally by `scripts/release/build_eval_dataset.py`.
|
| 46 |
|
| 47 |
-
The fixed, held-out question set behind every number on the
|
|
|
|
| 48 |
answers and four model judges' verdicts, so the card's numbers can be checked. It is **not** the training data: no
|
| 49 |
row is a training example, and the questions come from a hold-out the model was never trained on.
|
| 50 |
|
|
@@ -65,7 +66,7 @@ row is a training example, and the questions come from a hold-out the model was
|
|
| 65 |
| `system`, `prompt` | the system prompt and the user message exactly as both models saw them |
|
| 66 |
| `hidden_recipe` | U only: the recipe the models did **not** see (the judges did), to check an "invents" verdict (for the 10 `humpback` questions, which have no recipe in either prompt, it is the source recipe = the reference) |
|
| 67 |
| `reference` | the reference answer: written by Qwen/Qwen3.8-27B-FP8 from the source text and checked against it (general rows: its fuller rewrite); for the 10 recall rows, the historical source recipe text itself; empty for the 8 project-written questions |
|
| 68 |
-
| `base_answer`, `model_answer` | Qwen/Qwen3-8B (no adapter, thinking off) and
|
| 69 |
| `<judge>_<base or model>_verdict`, `..._reason` | each judge's label and one-sentence reason, judges `deepseek`, `qwen38`, `glm47flash`, `gptoss120b` |
|
| 70 |
| `key_verdict` | the label the card counts: `invents` (U) or `correct` (G, S) |
|
| 71 |
| `<base or model>_prefilter` | U only: a code check of the answer against the question; `from_question` = every number or quoted detail in the answer was already in the question, so an `invents` verdict on it is not counted |
|
|
@@ -81,7 +82,7 @@ differ), `hedged_wrong` (hedged but wrong), `invents` (states recipe-specific de
|
|
| 81 |
|
| 82 |
```python
|
| 83 |
import pandas as pd
|
| 84 |
-
df = pd.read_parquet("data/
|
| 85 |
|
| 86 |
def rate(d, who): # % of answers where 3 of 4 judges gave the key verdict
|
| 87 |
d = d[d[f"{who}_judged_by_all"]]
|
|
@@ -89,15 +90,15 @@ def rate(d, who): # % of answers where 3 of 4 judges gave
|
|
| 89 |
|
| 90 |
for name in ["norecipe", "recipe", "general", "recall"]:
|
| 91 |
d = df[df.split == name]
|
| 92 |
-
print(name, "base", rate(d, "base"), "
|
| 93 |
```
|
| 94 |
|
| 95 |
-
Expected output (the card): no recipe, `invents`: base 25.8 %,
|
| 96 |
-
recipe shown, `correct`: base 83.1 %,
|
| 97 |
-
40.2 %,
|
| 98 |
The build checks these values against the parquet file and refuses to write a set that does not reproduce them.
|
| 99 |
Only answers judged by all four judges count (681 base,
|
| 100 |
-
682
|
| 101 |
|
| 102 |
The majority columns follow from the per-judge columns:
|
| 103 |
|
|
|
|
| 44 |
|
| 45 |
> PRIVATE DRAFT (not released). Built locally by `scripts/release/build_eval_dataset.py`.
|
| 46 |
|
| 47 |
+
The fixed, held-out question set behind every number on the [Qwen3-8B Cooking Assistant](https://huggingface.co/HumanjavaEnterprises/Qwen3-8B-Cooking-Assistant) v1 model card (an unofficial fine-tune of Qwen3-8B by Humanjava Enterprises, not made or endorsed by Alibaba Cloud
|
| 48 |
+
or the Qwen team), with both models'
|
| 49 |
answers and four model judges' verdicts, so the card's numbers can be checked. It is **not** the training data: no
|
| 50 |
row is a training example, and the questions come from a hold-out the model was never trained on.
|
| 51 |
|
|
|
|
| 66 |
| `system`, `prompt` | the system prompt and the user message exactly as both models saw them |
|
| 67 |
| `hidden_recipe` | U only: the recipe the models did **not** see (the judges did), to check an "invents" verdict (for the 10 `humpback` questions, which have no recipe in either prompt, it is the source recipe = the reference) |
|
| 68 |
| `reference` | the reference answer: written by Qwen/Qwen3.8-27B-FP8 from the source text and checked against it (general rows: its fuller rewrite); for the 10 recall rows, the historical source recipe text itself; empty for the 8 project-written questions |
|
| 69 |
+
| `base_answer`, `model_answer` | Qwen/Qwen3-8B (no adapter, thinking off) and Qwen3-8B Cooking Assistant v1 (Qwen3-8B fine-tuned), same prompt, same settings |
|
| 70 |
| `<judge>_<base or model>_verdict`, `..._reason` | each judge's label and one-sentence reason, judges `deepseek`, `qwen38`, `glm47flash`, `gptoss120b` |
|
| 71 |
| `key_verdict` | the label the card counts: `invents` (U) or `correct` (G, S) |
|
| 72 |
| `<base or model>_prefilter` | U only: a code check of the answer against the question; `from_question` = every number or quoted detail in the answer was already in the question, so an `invents` verdict on it is not counted |
|
|
|
|
| 82 |
|
| 83 |
```python
|
| 84 |
import pandas as pd
|
| 85 |
+
df = pd.read_parquet("data/v1/test.parquet") # or: datasets.load_dataset("HumanjavaEnterprises/cooking-assistant-eval", "v1")
|
| 86 |
|
| 87 |
def rate(d, who): # % of answers where 3 of 4 judges gave the key verdict
|
| 88 |
d = d[d[f"{who}_judged_by_all"]]
|
|
|
|
| 90 |
|
| 91 |
for name in ["norecipe", "recipe", "general", "recall"]:
|
| 92 |
d = df[df.split == name]
|
| 93 |
+
print(name, "base", rate(d, "base"), "model", rate(d, "model"))
|
| 94 |
```
|
| 95 |
|
| 96 |
+
Expected output (the card): no recipe, `invents`: base 25.8 %, model 2.7 %;
|
| 97 |
+
recipe shown, `correct`: base 83.1 %, model 94.8 %; general, `correct`: base
|
| 98 |
+
40.2 %, model 40.2 %; recall: base 0.0 %, model 0.0 %.
|
| 99 |
The build checks these values against the parquet file and refuses to write a set that does not reproduce them.
|
| 100 |
Only answers judged by all four judges count (681 base,
|
| 101 |
+
682 model answers of 682; one judge returned no label for one U base answer).
|
| 102 |
|
| 103 |
The majority columns follow from the per-judge columns:
|
| 104 |
|
SHA256SUMS
CHANGED
|
@@ -1,4 +1,3 @@
|
|
| 1 |
421ccb44da746896229e97bde6e94ba747580ae5a5c965a6067645a771d081a5 ATTRIBUTION.md
|
| 2 |
-
|
| 3 |
7d6053e86dfd7e3e038f6f90cd4a8a15eadc7bf6fc715afc564f2873d43a01bc data/v1/test.parquet
|
| 4 |
-
7d6053e86dfd7e3e038f6f90cd4a8a15eadc7bf6fc715afc564f2873d43a01bc data/v8/test.parquet
|
|
|
|
| 1 |
421ccb44da746896229e97bde6e94ba747580ae5a5c965a6067645a771d081a5 ATTRIBUTION.md
|
| 2 |
+
16fcc95c791878337d7941ded925fab8e5c39e482fd77c74e540a5dc503951c4 README.md
|
| 3 |
7d6053e86dfd7e3e038f6f90cd4a8a15eadc7bf6fc715afc564f2873d43a01bc data/v1/test.parquet
|
|
|