Vveerrgg commited on
Commit
7d3512e
·
verified ·
1 Parent(s): 889080e

card: names the model and v1, non-affiliation line

Browse files
Files changed (2) hide show
  1. README.md +9 -8
  2. 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 Humanjava cooking LoRA (v8) model card, with both models'
 
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 the v8 model (Qwen3-8B + the v8 LoRA), same prompt, same settings |
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/v8/test.parquet") # or: datasets.load_dataset("HumanjavaEnterprises/cooking-assistant-eval", "v8")
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"), "v8", rate(d, "model"))
93
  ```
94
 
95
- Expected output (the card): no recipe, `invents`: base 25.8 %, v8 2.7 %;
96
- recipe shown, `correct`: base 83.1 %, v8 94.8 %; general, `correct`: base
97
- 40.2 %, v8 40.2 %; recall: base 0.0 %, v8 0.0 %.
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 v8 answers of 682; one judge returned no label for one U base answer).
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
- 3439a7dbade3d8d4e26e7334b96d50d598527bf0f2c2da5b7fe425e095e667ab README.md
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