Vveerrgg commited on
Commit
d463313
·
verified ·
1 Parent(s): 7d3512e

Add config glm-4-9b-v1 (GLM-4-9B Cooking Assistant v1); v1 unchanged

Browse files
Files changed (3) hide show
  1. README.md +82 -9
  2. SHA256SUMS +2 -1
  3. data/glm-4-9b-v1/test.parquet +3 -0
README.md CHANGED
@@ -26,7 +26,7 @@ task_categories:
26
  - question-answering
27
  size_categories:
28
  - n<1K
29
- pretty_name: Qwen3-8B Cooking Assistant evaluation set
30
  tags:
31
  - cooking
32
  - evaluation
@@ -38,11 +38,26 @@ configs:
38
  data_files:
39
  - split: test
40
  path: data/v1/test.parquet
 
 
 
 
41
  ---
42
 
43
- # Qwen3-8B Cooking Assistant evaluation set (config `v1`)
 
 
 
 
 
 
 
 
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'
@@ -58,7 +73,7 @@ row is a training example, and the questions come from a hold-out the model was
58
  | U | norecipe | 300 | the same G questions **without** the recipe; the model should ask for it or hedge, not invent amounts or steps |
59
  | S | general | 82 | general cooking questions that need no recipe |
60
 
61
- ## Columns
62
 
63
  | column | meaning |
64
  |---|---|
@@ -78,7 +93,7 @@ differ), `hedged_wrong` (hedged but wrong), `invents` (states recipe-specific de
78
  `correct_no_missing` (the question carried what was needed and the answer stays with it), `wrong`. G and S: `correct`,
79
  `partial`, `wrong`, `needless_ask`.
80
 
81
- ## Reproducing the card
82
 
83
  ```python
84
  import pandas as pd
@@ -114,7 +129,7 @@ def majority(r, who):
114
  assert all(df.apply(lambda r: majority(r, "model"), axis=1).dropna() == df.model_majority.dropna())
115
  ```
116
 
117
- ## The judges are models
118
 
119
  Every verdict here is a model's label, not a person's: DeepSeek-V4-Flash-0731 (MIT, unsloth UD-IQ2_M GGUF),
120
  Qwen3.8-27B-FP8 (Apache-2.0; it is also the model that wrote the training questions, so it is not independent of
@@ -123,7 +138,7 @@ question, the reference and, for U, the hidden recipe. Where they disagree, the
123
  majority measures agreement between four independent model families, not truth. Judges did not see which model
124
  wrote an answer.
125
 
126
- ## Sources and licence
127
 
128
  Text: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/), with per-row attribution. Rows made from
129
  Wikibooks pages are adaptations of CC BY-SA 4.0 material. 282 rows come from public-domain sources (`public_domain`
@@ -131,13 +146,71 @@ true: pre-1931 books from Project Gutenberg and the Internet Archive, and US-gov
131
  text stays public domain, and the licence covers our questions, answers, verdicts and the collection. Rows by licence:
132
  CC-BY-SA-4.0 400, LicenseRef-PD 216, LicenseRef-PDM-1.0 66. Full list: `ATTRIBUTION.md`. NHLBI does not endorse this dataset.
133
 
134
- ## Limitations
135
 
136
  - 682 rows: confidence intervals are wide (the card gives Wilson intervals), and the U and G rows share
137
  their questions.
138
  - Historical recipes use historical units and food-safety practice; they are sources, not advice.
139
  - Answers were generated once each (greedy decoding); another sample can differ.
140
 
141
- ## Files
142
 
143
  `data/v1/test.parquet` (config `v1`; each model version gets its own config, scored on the same questions), `ATTRIBUTION.md`, `SHA256SUMS`.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
26
  - question-answering
27
  size_categories:
28
  - n<1K
29
+ pretty_name: Cooking Assistant evaluation set
30
  tags:
31
  - cooking
32
  - evaluation
 
38
  data_files:
39
  - split: test
40
  path: data/v1/test.parquet
41
+ - config_name: glm-4-9b-v1
42
+ data_files:
43
+ - split: test
44
+ path: data/glm-4-9b-v1/test.parquet
45
  ---
46
 
47
+ # Cooking Assistant evaluation set
48
+
49
+ One fixed, held-out question set, used to score each Cooking Assistant model against its base model. One config per
50
+ model version, same questions:
51
+
52
+ | config | model | base model |
53
+ |---|---|---|
54
+ | `v1` (default) | [Qwen3-8B Cooking Assistant](https://huggingface.co/HumanjavaEnterprises/Qwen3-8B-Cooking-Assistant) v1 | Qwen/Qwen3-8B |
55
+ | `glm-4-9b-v1` | [GLM-4-9B Cooking Assistant](https://huggingface.co/HumanjavaEnterprises/GLM-4-9B-Cooking-Assistant) v1 | zai-org/GLM-4-9B-0414 |
56
 
57
+ Both are unofficial fine-tunes by Humanjava Enterprises, not made or endorsed by Alibaba Cloud, the Qwen team, Zhipu AI
58
+ or Z.ai. Each config carries the same columns except the judge and majority columns (see each section).
59
+
60
+ ## Config `v1`: Qwen3-8B Cooking Assistant v1
61
 
62
  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
63
  or the Qwen team), with both models'
 
73
  | U | norecipe | 300 | the same G questions **without** the recipe; the model should ask for it or hedge, not invent amounts or steps |
74
  | S | general | 82 | general cooking questions that need no recipe |
75
 
76
+ ### Columns
77
 
78
  | column | meaning |
79
  |---|---|
 
93
  `correct_no_missing` (the question carried what was needed and the answer stays with it), `wrong`. G and S: `correct`,
94
  `partial`, `wrong`, `needless_ask`.
95
 
96
+ ### Reproducing the card
97
 
98
  ```python
99
  import pandas as pd
 
129
  assert all(df.apply(lambda r: majority(r, "model"), axis=1).dropna() == df.model_majority.dropna())
130
  ```
131
 
132
+ ### The judges are models
133
 
134
  Every verdict here is a model's label, not a person's: DeepSeek-V4-Flash-0731 (MIT, unsloth UD-IQ2_M GGUF),
135
  Qwen3.8-27B-FP8 (Apache-2.0; it is also the model that wrote the training questions, so it is not independent of
 
138
  majority measures agreement between four independent model families, not truth. Judges did not see which model
139
  wrote an answer.
140
 
141
+ ### Sources and licence
142
 
143
  Text: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/), with per-row attribution. Rows made from
144
  Wikibooks pages are adaptations of CC BY-SA 4.0 material. 282 rows come from public-domain sources (`public_domain`
 
146
  text stays public domain, and the licence covers our questions, answers, verdicts and the collection. Rows by licence:
147
  CC-BY-SA-4.0 400, LicenseRef-PD 216, LicenseRef-PDM-1.0 66. Full list: `ATTRIBUTION.md`. NHLBI does not endorse this dataset.
148
 
149
+ ### Limitations
150
 
151
  - 682 rows: confidence intervals are wide (the card gives Wilson intervals), and the U and G rows share
152
  their questions.
153
  - Historical recipes use historical units and food-safety practice; they are sources, not advice.
154
  - Answers were generated once each (greedy decoding); another sample can differ.
155
 
156
+ ### Files
157
 
158
  `data/v1/test.parquet` (config `v1`; each model version gets its own config, scored on the same questions), `ATTRIBUTION.md`, `SHA256SUMS`.
159
+
160
+ ## Config `glm-4-9b-v1`: GLM-4-9B Cooking Assistant v1
161
+
162
+ The same 682 questions, answered by base GLM-4-9B-0414 and by
163
+ [GLM-4-9B Cooking Assistant](https://huggingface.co/HumanjavaEnterprises/GLM-4-9B-Cooking-Assistant) v1 (an unofficial
164
+ fine-tune of GLM-4-9B-0414 by Humanjava Enterprises, not made or endorsed by Zhipu AI / Z.ai): the numbers behind that
165
+ model card. As for `v1`, no row is a training example. The columns are those of `v1`, except:
166
+
167
+ | column | meaning in `glm-4-9b-v1` |
168
+ |---|---|
169
+ | `base_answer`, `model_answer` | zai-org/GLM-4-9B-0414 (no adapter) and GLM-4-9B Cooking Assistant v1 (GLM-4-9B-0414 fine-tuned), same prompt, same settings, both from full-precision weights |
170
+ | `<judge>_<base or model>_verdict`, `..._reason` | judges `deepseek`, `qwen38`, `gptoss120b`, `seedoss36b` (no GLM-family judge: Seed-OSS-36B-Instruct takes the place of GLM-4.7-Flash) |
171
+ | `<base or model>_judged_by_all`, `<base or model>_majority` | the headline: the three judges from outside both the Qwen and the GLM families (`deepseek`, `gptoss120b`, `seedoss36b`) all returned a label; at least 2 of 3 gave `key_verdict` (after the code check) |
172
+ | `<base or model>_judged_by_all4`, `<base or model>_majority4` | all four judges returned a label; at least 3 of 4 gave `key_verdict`. Secondary: `qwen38` is a Qwen-family model and wrote the training questions |
173
+
174
+ ### Reproducing the card
175
+
176
+ ```python
177
+ import pandas as pd
178
+ df = pd.read_parquet("data/glm-4-9b-v1/test.parquet") # or: datasets.load_dataset("HumanjavaEnterprises/cooking-assistant-eval", "glm-4-9b-v1")
179
+
180
+ def rate(d, who, sfx=""): # sfx="": 2 of 3 family-neutral judges; sfx="4": 3 of 4 judges
181
+ d = d[d[f"{who}_judged_by_all{sfx}"]]
182
+ return round(100 * d[f"{who}_majority{sfx}"].mean(), 1)
183
+
184
+ for name in ["norecipe", "recipe", "general", "recall"]:
185
+ d = df[df.split == name]
186
+ print(name, "base", rate(d, "base"), "model", rate(d, "model"), "| 4 judges:", rate(d, "base", "4"), rate(d, "model", "4"))
187
+ ```
188
+
189
+ Expected output (the card), headline (2 of 3): no recipe, `invents`: base 40.1 %, model
190
+ 9.7 %; recipe shown, `correct`: base 83.1 %, model 96.2 %; general,
191
+ `correct`: base 56.1 %, model 46.3 %; recall: base 20.0 %, model
192
+ 10.0 %. All four judges (3 of 4): no recipe 30.1 % / 4.7 %, recipe
193
+ shown 78.6 % / 95.5 %, general 40.2 % / 36.6 %, recall
194
+ 0.0 % / 10.0 %. Only answers judged by every judge of that table count (headline:
195
+ 681 base, 681 model answers of 682; four judges:
196
+ 681 base, 681 model). The build checks these values against the
197
+ parquet file and refuses to write a set that does not reproduce them.
198
+
199
+ ```python
200
+ def majority(r, who, judges):
201
+ v = [r[f"{j}_{who}_verdict"] for j in judges]
202
+ if any(x is None for x in v):
203
+ return None
204
+ hit = [x == r.key_verdict and not (r.key_verdict == "invents" and r[f"{who}_prefilter"] == "from_question")
205
+ for x in v]
206
+ return sum(hit) > len(v) / 2
207
+ headline = ["deepseek", "gptoss120b", "seedoss36b"]
208
+ assert all(df.apply(lambda r: majority(r, "model", headline), axis=1).dropna() == df.model_majority.dropna())
209
+ assert all(df.apply(lambda r: majority(r, "model", headline + ["qwen38"]), axis=1).dropna() == df.model_majority4.dropna())
210
+ ```
211
+
212
+ The judges: DeepSeek-V4-Flash-0731 (MIT, unsloth UD-IQ2_M GGUF), gpt-oss-120b (Apache-2.0), Seed-OSS-36B-Instruct by
213
+ ByteDance Seed (Apache-2.0) and Qwen3.8-27B-FP8 (Apache-2.0), all run locally from open weights, blind to which model
214
+ wrote an answer. As in `v1`, the verdicts are model labels, not truth.
215
+
216
+ `data/glm-4-9b-v1/test.parquet` (config `glm-4-9b-v1`).
SHA256SUMS CHANGED
@@ -1,3 +1,4 @@
1
  421ccb44da746896229e97bde6e94ba747580ae5a5c965a6067645a771d081a5 ATTRIBUTION.md
2
- 16fcc95c791878337d7941ded925fab8e5c39e482fd77c74e540a5dc503951c4 README.md
 
3
  7d6053e86dfd7e3e038f6f90cd4a8a15eadc7bf6fc715afc564f2873d43a01bc data/v1/test.parquet
 
1
  421ccb44da746896229e97bde6e94ba747580ae5a5c965a6067645a771d081a5 ATTRIBUTION.md
2
+ e665f8a2315216a20ee48fa4c9251df01d28ec3bb09de0ca87574e23248fe268 README.md
3
+ c4fb809e9d31b5c3fb5c400e33fdba2120bedaf2134eb0f76b58ce47d7b2811f data/glm-4-9b-v1/test.parquet
4
  7d6053e86dfd7e3e038f6f90cd4a8a15eadc7bf6fc715afc564f2873d43a01bc data/v1/test.parquet
data/glm-4-9b-v1/test.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c4fb809e9d31b5c3fb5c400e33fdba2120bedaf2134eb0f76b58ce47d7b2811f
3
+ size 771622