Vveerrgg commited on
Commit
2b0c781
·
verified ·
1 Parent(s): 5f1d3cf

evaluation set behind the v8 card numbers

Browse files
Files changed (4) hide show
  1. ATTRIBUTION.md +63 -0
  2. README.md +141 -0
  3. SHA256SUMS +3 -0
  4. data/test.parquet +3 -0
ATTRIBUTION.md ADDED
@@ -0,0 +1,63 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Attribution
2
+
3
+ Every row names its source in its own fields (`source_title`, `source_url`, `source_licence`, `source_attribution`, `source_author`, and for Wikibooks `source_history_url` + `source_revision_id`). This file groups them.
4
+
5
+ ## Wikibooks (CC BY-SA 4.0)
6
+
7
+ Recipe text from Wikibooks cookbooks, licensed CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/). Authors: each page's history (`source_history_url`); `source_revision_id` is the revision read. Changes: wiki markup converted to plain text; passages may be cut; questions and answers about the text were written by the models named in the card.
8
+
9
+ | Wikibooks | pages | rows |
10
+ |---|---|---|
11
+ | ar.wikibooks.org | 3 | 9 |
12
+ | bn.wikibooks.org | 2 | 15 |
13
+ | de.wikibooks.org | 5 | 32 |
14
+ | en.wikibooks.org | 4 | 31 |
15
+ | es.wikibooks.org | 5 | 22 |
16
+ | fa.wikibooks.org | 4 | 25 |
17
+ | fr.wikibooks.org | 5 | 16 |
18
+ | he.wikibooks.org | 1 | 6 |
19
+ | id.wikibooks.org | 5 | 32 |
20
+ | it.wikibooks.org | 2 | 12 |
21
+ | ja.wikibooks.org | 6 | 41 |
22
+ | ms.wikibooks.org | 1 | 4 |
23
+ | nl.wikibooks.org | 4 | 34 |
24
+ | pl.wikibooks.org | 3 | 14 |
25
+ | pt.wikibooks.org | 6 | 38 |
26
+ | sv.wikibooks.org | 2 | 21 |
27
+ | th.wikibooks.org | 1 | 2 |
28
+ | tl.wikibooks.org | 1 | 8 |
29
+ | tr.wikibooks.org | 1 | 11 |
30
+ | vi.wikibooks.org | 2 | 15 |
31
+ | zh.wikibooks.org | 1 | 4 |
32
+
33
+ ## Public-domain books
34
+
35
+ Pre-1931 cookbooks; `pd_basis` gives the public-domain basis of each. Internet Archive scans were re-OCRed on our machine (allenai/olmOCR-2-7B-1025-FP8, Apache-2.0).
36
+
37
+ | work | author | URL | public-domain basis | rows |
38
+ |---|---|---|---|---|
39
+ | Canadian Housewife's Manual of Cookery (1861) | | https://archive.org/details/cihm_64026 | Published in 1861; public domain in the US | 9 |
40
+ | Cozinheiro nacional (c. 1870s) | Rodrigo | https://archive.org/details/cozinheiro-nacional | Public Domain Mark on the source; published by 1889; public domain in the US | 6 |
41
+ | Five Roses Cook Book (1913) | Lake of the Woods Milling Company | https://archive.org/details/McGillLibrary-rbsc_five-roses-cook-book_TX769F581913-23364 | Published in 1913; public domain in the US | 8 |
42
+ | Galván Rivera, Nuevo cocinero mexicano en forma de diccionario (1845) | Mariano Galván Rivera | https://archive.org/details/bub_gb_NdQqAAAAYAAJ | Public Domain Mark on the source; published in 1845; public domain in the US | 60 |
43
+ | Nouvelle cuisinière canadienne (Montréal 1879) | | https://archive.org/details/cihm_13725 | Published in 1879; public domain in the US | 4 |
44
+ | The Home Cook Book (Toronto 1877) | compiled from recipes contributed by ladies of Toronto and other cities and towns. | https://archive.org/details/cihm_95365 | Published in 1877; public domain in the US | 26 |
45
+ | Artusi, La scienza in cucina (1891) | Pellegrino Artusi | https://www.gutenberg.org/ebooks/59047 | Published in 1922; public domain in the US; author died in 1911 | 5 |
46
+ | Beeton, Book of Household Management (1861) | Mrs. Beeton | https://www.gutenberg.org/ebooks/10136 | Published in 1861; public domain in the US; author died in 1865 | 48 |
47
+ | Borgh, Zeniths Kokbok | Anna Borgh | https://www.gutenberg.org/ebooks/57621 | Published in 1915; public domain in the US | 6 |
48
+ | Calleja, La Mejor Cocinera | Calleja | https://www.gutenberg.org/ebooks/8870 | Published in 1915; public domain in the US; author died in 1915 | 12 |
49
+ | Cocina moderna (anon.) | Anonymous | https://www.gutenberg.org/ebooks/45770 | Published in 1875; public domain in the US | 22 |
50
+ | Davidis, Keukenboek (Dutch tr.) | Henriette Davidis | https://www.gutenberg.org/ebooks/37790 | Translation published in 1873 of a work first published in 1867; public domain in the US | 13 |
51
+ | Escoffier, A Guide to Modern Cookery (1907 EN, anon. tr.) | A. Escoffier | https://www.gutenberg.org/ebooks/71395 | Published in 1909; public domain in the US; author died in 1935 | 25 |
52
+ | Gillette & Ziemann, The White House Cook Book (1887) | F. L. Gillette; Hugo Ziemann | https://www.gutenberg.org/ebooks/13923 | Published in 1913; public domain in the US; author died in 1904 | 10 |
53
+ | Heyl, Volks-Kochbuch | Hedwig Heyl | https://www.gutenberg.org/ebooks/13921 | Published in 1905; public domain in the US; author died in 1934 | 7 |
54
+ | Lazzari-Turco, Il piccolo focolare | Giulia Lazzari-Turco | https://www.gutenberg.org/ebooks/65179 | Published in 1921; public domain in the US; author died in 1912 | 11 |
55
+ | Rossi, La vera cuciniera genovese | Emanuele Rossi | https://www.gutenberg.org/ebooks/51857 | Published by 1865; public domain in the US | 10 |
56
+
57
+ ## US federal government (public domain)
58
+
59
+ NHLBI (NIH) healthy-eating recipes: 0 pages, 0 rows (17 USC 105). NHLBI does not endorse this dataset.
60
+
61
+ ## Project-written questions
62
+
63
+ 8 general questions (`open0`..`open7`) were written by Humanjava for this evaluation and use no source text.
README.md ADDED
@@ -0,0 +1,141 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: cc-by-sa-4.0
3
+ language:
4
+ - ar
5
+ - bn
6
+ - de
7
+ - en
8
+ - es
9
+ - fa
10
+ - fr
11
+ - he
12
+ - id
13
+ - it
14
+ - ja
15
+ - ms
16
+ - nl
17
+ - pl
18
+ - pt
19
+ - sv
20
+ - th
21
+ - tl
22
+ - tr
23
+ - vi
24
+ - zh
25
+ task_categories:
26
+ - question-answering
27
+ size_categories:
28
+ - n<1K
29
+ pretty_name: Cooking LoRA v8 evaluation set
30
+ tags:
31
+ - cooking
32
+ - evaluation
33
+ - llm-as-judge
34
+ - multilingual
35
+ configs:
36
+ - config_name: default
37
+ data_files:
38
+ - split: test
39
+ path: data/test.parquet
40
+ ---
41
+
42
+ # Cooking LoRA v8 evaluation set
43
+
44
+ > PRIVATE DRAFT (not released). Built locally by `scripts/release/build_eval_dataset.py`.
45
+
46
+ The fixed, held-out question set behind every number on the Humanjava cooking LoRA (v8) model card, with both models'
47
+ answers and four model judges' verdicts, so the card's numbers can be checked. It is **not** the training data: no
48
+ row is a training example, and the questions come from a hold-out the model was never trained on.
49
+
50
+ 682 rows, one per question and setting:
51
+
52
+ | set | split | rows | what it tests |
53
+ |---|---|---|---|
54
+ | G | recipe | 290 | the recipe is in the prompt; does the model read it correctly? |
55
+ | G | recall | 10 | "how do you make X?" where the reference is a historical recipe text; wrong by design for both models (no one recalls an 1800s recipe word for word) |
56
+ | U | norecipe | 300 | the same G questions **without** the recipe; the model should ask for it or hedge, not invent amounts or steps |
57
+ | S | general | 82 | general cooking questions that need no recipe |
58
+
59
+ ## Columns
60
+
61
+ | column | meaning |
62
+ |---|---|
63
+ | `id`, `set`, `kind`, `split`, `type`, `lang` | question id (shared by its G and U rows), set, plain-words setting, scoring split, item type, language |
64
+ | `system`, `prompt` | the system prompt and the user message exactly as both models saw them |
65
+ | `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) |
66
+ | `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 |
67
+ | `base_answer`, `model_answer` | Qwen/Qwen3-8B (no adapter, thinking off) and the v8 model (Qwen3-8B + the v8 LoRA), same prompt, same settings |
68
+ | `<judge>_<base or model>_verdict`, `..._reason` | each judge's label and one-sentence reason, judges `deepseek`, `qwen38`, `glm47flash`, `gptoss120b` |
69
+ | `key_verdict` | the label the card counts: `invents` (U) or `correct` (G, S) |
70
+ | `<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 |
71
+ | `<base or model>_judged_by_all`, `<base or model>_majority` | all four judges returned a label; at least 3 of 4 gave `key_verdict` (after the code check) |
72
+ | `source_*`, `pd_basis`, `public_domain` | the source the question was made from: title, book, URL, Wikibooks history + revision, author, licence, ready-to-use attribution line |
73
+
74
+ Labels, in plain words. U: `asks` (asks for the recipe), `hedged_useful` (a general answer that says the recipe may
75
+ differ), `hedged_wrong` (hedged but wrong), `invents` (states recipe-specific details it could not know),
76
+ `correct_no_missing` (the question carried what was needed and the answer stays with it), `wrong`. G and S: `correct`,
77
+ `partial`, `wrong`, `needless_ask`.
78
+
79
+ ## Reproducing the card
80
+
81
+ ```python
82
+ import pandas as pd
83
+ df = pd.read_parquet("data/test.parquet")
84
+
85
+ def rate(d, who): # % of answers where 3 of 4 judges gave the key verdict
86
+ d = d[d[f"{who}_judged_by_all"]]
87
+ return round(100 * d[f"{who}_majority"].mean(), 1)
88
+
89
+ for name in ["norecipe", "recipe", "general", "recall"]:
90
+ d = df[df.split == name]
91
+ print(name, "base", rate(d, "base"), "v8", rate(d, "model"))
92
+ ```
93
+
94
+ Expected output (the card): no recipe, `invents`: base 25.8 %, v8 2.7 %;
95
+ recipe shown, `correct`: base 83.1 %, v8 94.8 %; general, `correct`: base
96
+ 40.2 %, v8 40.2 %; recall: base 0.0 %, v8 0.0 %.
97
+ The build checks these values against the parquet file and refuses to write a set that does not reproduce them.
98
+ Only answers judged by all four judges count (681 base,
99
+ 682 v8 answers of 682; one judge returned no label for one U base answer).
100
+
101
+ The majority columns follow from the per-judge columns:
102
+
103
+ ```python
104
+ judges = ["deepseek", "qwen38", "glm47flash", "gptoss120b"]
105
+ def majority(r, who):
106
+ v = [r[f"{j}_{who}_verdict"] for j in judges]
107
+ if any(x is None for x in v):
108
+ return None
109
+ hit = [x == r.key_verdict and not (r.key_verdict == "invents" and r[f"{who}_prefilter"] == "from_question")
110
+ for x in v]
111
+ return sum(hit) >= 3
112
+ assert all(df.apply(lambda r: majority(r, "model"), axis=1).dropna() == df.model_majority.dropna())
113
+ ```
114
+
115
+ ## The judges are models
116
+
117
+ Every verdict here is a model's label, not a person's: DeepSeek-V4-Flash-0731 (MIT, unsloth UD-IQ2_M GGUF),
118
+ Qwen3.8-27B-FP8 (Apache-2.0; it is also the model that wrote the training questions, so it is not independent of
119
+ the data), GLM-4.7-Flash (MIT, FP8) and gpt-oss-120b (Apache-2.0), all run locally from open weights. They saw the
120
+ question, the reference and, for U, the hidden recipe. Where they disagree, the reasons usually show why. The
121
+ majority measures agreement between four independent model families, not truth. Judges did not see which model
122
+ wrote an answer.
123
+
124
+ ## Sources and licence
125
+
126
+ Text: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/), with per-row attribution. Rows made from
127
+ Wikibooks pages are adaptations of CC BY-SA 4.0 material. 282 rows come from public-domain sources (`public_domain`
128
+ true: pre-1931 books from Project Gutenberg and the Internet Archive, and US-government NHLBI pages); their source
129
+ text stays public domain, and the licence covers our questions, answers, verdicts and the collection. Rows by licence:
130
+ CC-BY-SA-4.0 400, LicenseRef-PD 216, LicenseRef-PDM-1.0 66. Full list: `ATTRIBUTION.md`. NHLBI does not endorse this dataset.
131
+
132
+ ## Limitations
133
+
134
+ - 682 rows: confidence intervals are wide (the card gives Wilson intervals), and the U and G rows share
135
+ their questions.
136
+ - Historical recipes use historical units and food-safety practice; they are sources, not advice.
137
+ - Answers were generated once each (greedy decoding); another sample can differ.
138
+
139
+ ## Files
140
+
141
+ `data/test.parquet`, `ATTRIBUTION.md`, `SHA256SUMS`.
SHA256SUMS ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ 421ccb44da746896229e97bde6e94ba747580ae5a5c965a6067645a771d081a5 ATTRIBUTION.md
2
+ c3408153e4eb751cd127b994feb5544a7fed75bc7552760ad72fc6f778ab7fa0 README.md
3
+ 7d6053e86dfd7e3e038f6f90cd4a8a15eadc7bf6fc715afc564f2873d43a01bc data/test.parquet
data/test.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7d6053e86dfd7e3e038f6f90cd4a8a15eadc7bf6fc715afc564f2873d43a01bc
3
+ size 766766