HumanjavaEnterprises/cooking-assistant-eval

Qwen3-8B Cooking Assistant evaluation set (config v1)

PRIVATE DRAFT (not released). Built locally by scripts/release/build_eval_dataset.py.

The fixed, held-out question set behind every number on the Qwen3-8B Cooking Assistant v1 model card (an unofficial fine-tune of Qwen3-8B by Humanjava Enterprises, not made or endorsed by Alibaba Cloud or the Qwen team), with both models' answers and four model judges' verdicts, so the card's numbers can be checked. It is not the training data: no row is a training example, and the questions come from a hold-out the model was never trained on.

682 rows, one per question and setting:

set	split	rows	what it tests
G	recipe	290	the recipe is in the prompt; does the model read it correctly?
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)
U	norecipe	300	the same G questions without the recipe; the model should ask for it or hedge, not invent amounts or steps
S	general	82	general cooking questions that need no recipe

Columns

column	meaning
id, set, kind, split, type, lang	question id (shared by its G and U rows), set, plain-words setting, scoring split, item type, language
system, prompt	the system prompt and the user message exactly as both models saw them
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)
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
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
<judge>_<base or model>_verdict, ..._reason	each judge's label and one-sentence reason, judges deepseek, qwen38, glm47flash, gptoss120b
key_verdict	the label the card counts: invents (U) or correct (G, S)
<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
<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)
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

Labels, in plain words. U: asks (asks for the recipe), hedged_useful (a general answer that says the recipe may differ), hedged_wrong (hedged but wrong), invents (states recipe-specific details it could not know), correct_no_missing (the question carried what was needed and the answer stays with it), wrong. G and S: correct, partial, wrong, needless_ask.

Reproducing the card

import pandas as pd
df = pd.read_parquet("data/v1/test.parquet")   # or: datasets.load_dataset("HumanjavaEnterprises/cooking-assistant-eval", "v1")

def rate(d, who):                       # % of answers where 3 of 4 judges gave the key verdict
    d = d[d[f"{who}_judged_by_all"]]
    return round(100 * d[f"{who}_majority"].mean(), 1)

for name in ["norecipe", "recipe", "general", "recall"]:
    d = df[df.split == name]
    print(name, "base", rate(d, "base"), "model", rate(d, "model"))

Expected output (the card): no recipe, invents: base 25.8 %, model 2.7 %; recipe shown, correct: base 83.1 %, model 94.8 %; general, correct: base 40.2 %, model 40.2 %; recall: base 0.0 %, model 0.0 %. The build checks these values against the parquet file and refuses to write a set that does not reproduce them. Only answers judged by all four judges count (681 base, 682 model answers of 682; one judge returned no label for one U base answer).

The majority columns follow from the per-judge columns:

judges = ["deepseek", "qwen38", "glm47flash", "gptoss120b"]
def majority(r, who):
    v = [r[f"{j}_{who}_verdict"] for j in judges]
    if any(x is None for x in v):
        return None
    hit = [x == r.key_verdict and not (r.key_verdict == "invents" and r[f"{who}_prefilter"] == "from_question")
           for x in v]
    return sum(hit) >= 3
assert all(df.apply(lambda r: majority(r, "model"), axis=1).dropna() == df.model_majority.dropna())

The judges are models

Every verdict here is a model's label, not a person's: DeepSeek-V4-Flash-0731 (MIT, unsloth UD-IQ2_M GGUF), Qwen3.8-27B-FP8 (Apache-2.0; it is also the model that wrote the training questions, so it is not independent of the data), GLM-4.7-Flash (MIT, FP8) and gpt-oss-120b (Apache-2.0), all run locally from open weights. They saw the question, the reference and, for U, the hidden recipe. Where they disagree, the reasons usually show why. The majority measures agreement between four independent model families, not truth. Judges did not see which model wrote an answer.

Sources and licence

Text: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/), with per-row attribution. Rows made from Wikibooks pages are adaptations of CC BY-SA 4.0 material. 282 rows come from public-domain sources (public_domain true: pre-1931 books from Project Gutenberg and the Internet Archive, and US-government NHLBI pages); their source text stays public domain, and the licence covers our questions, answers, verdicts and the collection. Rows by licence: CC-BY-SA-4.0 400, LicenseRef-PD 216, LicenseRef-PDM-1.0 66. Full list: ATTRIBUTION.md. NHLBI does not endorse this dataset.

Limitations

• 682 rows: confidence intervals are wide (the card gives Wilson intervals), and the U and G rows share their questions.
• Historical recipes use historical units and food-safety practice; they are sources, not advice.
• Answers were generated once each (greedy decoding); another sample can differ.

Files

data/v1/test.parquet (config v1; each model version gets its own config, scored on the same questions), ATTRIBUTION.md, SHA256SUMS.