HumanjavaEnterprises/Qwen3-8B-Cooking-Assistant

Qwen3-8B Cooking Assistant

A cooking assistant that reads the recipe you give it, and asks for the recipe instead of inventing details when it can't see one. On questions that need a recipe it wasn't shown, it invents details 2.7% of the time vs 25.8% for base Qwen3-8B; given the recipe, it answers correctly 94.8% of the time vs 83.1% (four open judges, 3 of 4 agreeing). General cooking answers are level with the base. 26 languages; runs on a 6 GB GPU (Q4_K_M).

Unofficial fine-tune of Qwen3-8B by Humanjava Enterprises; not made or endorsed by Alibaba Cloud or the Qwen team.

Full-precision (bf16) safetensors for Transformers and vLLM. For llama.cpp, Ollama and LM Studio: HumanjavaEnterprises/Qwen3-8B-Cooking-Assistant-GGUF.

It does not memorise recipes: give it the recipe (paste it, or have your app look it up) and it will read it.

Files

The two training stages (supervised LoRA, then a preference LoRA on top) are merged into one bf16 model, the same architecture and tokenizer as Qwen3-8B (~16 GB). The GGUF files in the sibling repo are converted from these weights.

Quick start

# vLLM
vllm serve HumanjavaEnterprises/Qwen3-8B-Cooking-Assistant --max-model-len 8192
# then send requests with "chat_template_kwargs": {"enable_thinking": false}

# Transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "HumanjavaEnterprises/Qwen3-8B-Cooking-Assistant"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="auto", device_map="auto")
messages = [{"role": "system", "content": "You are a cooking assistant. You answer questions about dishes, ingredients, techniques and food safety. When you are not sure, you say so."},
            {"role": "user", "content": "If I double the recipe, how much pepper do I need?"}]
ids = tok.apply_chat_template(messages, add_generation_prompt=True, enable_thinking=False, return_tensors="pt").to(model.device)
print(tok.decode(model.generate(ids, max_new_tokens=400)[0][ids.shape[1]:], skip_special_tokens=True))

Versions

The model is one product with dated versions. Each version is a git tag on this repo, so you can pin one with revision="v1" (Transformers / huggingface_hub) or ollama run vveerrgg/qwen3-8b-cooking-assistant:v1.

Version	Date	Invents (no recipe, 3 of 4 judges)	Correct (recipe shown, 3 of 4 judges)	Notes
v1 (current)	2026-11-12	2.7% (base 25.8%)	94.8% (base 83.1%)	first public release

Prompt format

Qwen3 chat template, thinking off (enable_thinking: false, or /no_think in the user turn). Trained with this system prompt, which works best:

You are a cooking assistant. You answer questions about dishes, ingredients, techniques and food safety. When you are not sure, you say so.

To ask about a specific recipe, put the recipe text in the user turn, then the question.

What it does

The fine-tune teaches behaviour, not facts:

• Reads the recipe you give it. Paste a recipe and ask about it; it answers from that text.
• Asks when the recipe is missing. "If I double the recipe, how much pepper?" with no recipe gets a request for the original quantity, not an invented number.
• Answers general cooking questions in full: food safety, substitutions, why a dish failed.

The intended setup is LoRA for behaviour, retrieval for facts, tools for numbers: an app looks up the recipe and hands it to the model; scaling and unit conversion are calculated, not guessed.

It is also a worked example of a small custom model built carefully: open weights, open or public-domain sources only, every training record traceable to its source, and the checks published along with the failures they caught.

Results

All cooking scores are on a fixed held-out set of 682 questions in 26 languages, never used in training: 290 with the recipe shown, 10 recall questions with no recipe ("how do you make X?"), 300 that need a recipe the model is not shown, and 82 general questions. Answers from base Qwen3-8B (thinking off) and this model are compared by an open-weights judge from a different model family (DeepSeek-V4-Flash, MIT licence), blind, in both orders. 95% confidence intervals in brackets.

	Qwen3-8B	this model
Invents recipe details when no recipe is shown	30.0% [25.1, 35.4]	9.0% [6.3, 12.8]
same, after a rule check that removes judge false alarms	27.7%	4.0% [2.3, 6.9]
Correct when the recipe is shown (290 questions)	87.6%	96.2%
Preferred by the judge, all 682 questions	—	0.610 [0.581, 0.639] (0.50 = tie)
Preferred by the judge, general questions only	—	0.460 [0.369, 0.549] (level with base)

The drop in invented details is significant (McNemar p < 1e-10; 78 questions where only the base invents, 15 where only this model does).

Four judges. To check the judge, three more open-weights judges from different families scored the same answers one at a time: Qwen3.8-27B (Apache-2.0), GLM-4.7-Flash (MIT) and gpt-oss-120b (Apache-2.0). With the same rule check applied to all four, the share of no-recipe answers that invent recipe details:

counted as "invents" when	Qwen3-8B	this model
at least 3 of 4 judges say so	25.8% [21.1, 31.0]	2.7% [1.4, 5.2]
all four say so	19.1%	1.0%
any one says so	49.2%	18.3%

The same four judges on the questions where the recipe is shown and the model has to read it (290 questions; 3 of 4 must agree): Qwen3-8B correct 83.1% [78.4, 87.0], this model 94.8% [91.6, 96.8] (40 questions only this model gets right, 6 only the base does, McNemar p < 1e-6). Two other kinds of question in the set are reported apart: 82 general questions graded against a reference answer (40.2% for both models; the judges disagree widely there) and 10 "how do you make X?" questions with no recipe whose reference is the original 1800s text (0% for both: the model is built not to recall historical recipes from memory).

Every judge, and every way of combining them, shows a significant drop (3 of 4: 72 questions where only the base invents, 3 where only this model does, McNemar p < 1e-17). The judges agree moderately on base answers (Fleiss kappa 0.63). GLM-4.7-Flash flags more of this model's answers than the others; reading them, most repeat or restate a line the user quoted, which the rubric says is not inventing, and the rule check now clears those for every judge. Qwen3.8-27B also labelled this model's preference training data, so its read of this model's answers is not fully independent. Model judges also tend to make the same mistakes as each other, so four agreeing is less than four independent checks. Most of the remaining "invents" cases are borderline: a quantity that appears in the question itself, or a common default ("a pinch of salt") offered with a hedge.

Quantised files were scored on the same 682 questions against the full-precision model (0.50 = identical quality):

	Q4_K_M	Q6_K	Q8_0
Judge preference vs full precision	0.492 [0.475, 0.509]	0.501 [0.485, 0.517]	0.481 [0.466, 0.494]
Correct when the recipe is shown, 290 questions (full precision on that run)	97.2% (96.6%)	96.6% (96.2%)	96.9% (96.2%)
Invents with no recipe (full precision on that run)	9.3% (7.3%)	6.3% (6.3%)	6.7% (6.7%)
Invents with no recipe, 3 of 4 judges (full precision)	4.7% (3.3%)	3.0% (3.3%)	3.3% (2.7%)
Correct when the recipe is shown, 3 of 4 judges (full precision)	95.5% (94.1%)	95.5% (94.1%)	95.2% (94.8%)

With all four judges, no quantised file differs significantly from full precision (McNemar on the 3-of-4 verdict: Q4_K_M p = 0.34, Q6_K p = 1.0, Q8_0 p = 0.69 for inventing; p ≥ 0.22 for correctness with the recipe shown). On Q4_K_M two of the four judges on their own (GLM-4.7-Flash, gpt-oss-120b) do see slightly more invented details (p = 0.02 and 0.04); pick Q6_K or Q8_0 if that matters.

With the DeepSeek judge alone, no quantisation changes the invents rate significantly (McNemar p ≥ 0.29). Q8_0's small dip comes from the 82 general questions (0.418) and is most likely judge noise on a small set; Q4_K_M and Q6_K, which remove more bits, sit at parity.

Memorisation. 1,000 CC BY-SA training documents (Wikibooks) were probed 12,000 ways (prefixes of 32 to 128 tokens at three offsets, raw and chat prompts, greedy decoding). The model never continues a document verbatim (no 50-token match, longest overlap 67 characters vs the base model's 52), and overlap with the true text is no higher than base Qwen3-8B's (ROUGE-L 0.169 vs 0.177).

General ability (lm-eval 0.4.13, thinking off; MMLU-Pro on 100 questions per subject, so ±1.3 points is noise):

	IFEval	GSM8K	TruthfulQA mc2	MMLU-Pro	MGSM (8 lang)	Belebele (21 lang)
Qwen3-8B	0.817	0.807	0.532	0.611	0.341	0.846
this model	0.800	0.866	0.542	0.591	0.401	0.863

MMLU-Pro is 2.0 points down, the largest loss. The GSM8K and MGSM gains most likely reflect cleaner final-answer formatting learned from the chat data, not better maths. Belebele (reading comprehension) is within 2 points of the base in all 21 languages.

Refusals (XSTest, 250 safe and 200 unsafe prompts). Without a system prompt it behaves like the base (refuses 5.2% of safe prompts vs 4.4%). With the cooking system prompt, the base refuses 38.4% of safe prompts as off-topic; this model refuses 18.4% and answers or redirects the rest. We read every unsafe prompt the base refused and this model did not: all were deflections or explicit rebuttals, none complied.

How it was built

Sources	Cooking texts in 26 languages from openly licensed and public-domain sources only: the Wikibooks Cookbook (CC BY-SA 4.0), public-domain cookbooks from Project Gutenberg and the Internet Archive (scans re-read with olmOCR-2), and NHLBI recipes (US government, public domain). Every record carries its source URL, licence and public-domain basis. About a quarter of the training records are English
Excluded	Traditional and Indigenous knowledge, sacred material, outsider-authored books about other cultures, dog-meat recipes, anything without a verified licence
Models used	Open-weights models only, all Apache-2.0 or MIT. No output from any closed or API model is in the training, evaluation or seed data
Training	Two LoRA stages on Qwen3-8B, each merged: a supervised stage that teaches reading a shown recipe, asking when there is none, and full general answers; then a preference stage that favours asking over inventing and correct over wrong. General chat data is mixed in to limit forgetting
Tooling	LLaMA-Factory, vLLM, llama.cpp

The evaluation set behind every number on this card (all 682 questions, both models' answers, every judge's verdict, and per-row source credits) is published so the results can be checked: HumanjavaEnterprises/cooking-assistant-eval.

What went wrong along the way

• The first model invented recipe details. Questions had been written while looking at a recipe, then trained without it, so the model learned to answer as if it could see one. Benchmarks did not catch it; reading the answers did. Fix: the recipe goes in the prompt, plus examples of asking when it is missing.
• Then the general answers got thin. Teaching it to ask made general answers short (median 440 characters vs the base's 1,223), and lowering the adapter strength brought depth back but lost the asking. Fix: let the base model rewrite general answers in its own words, then a preference stage on the model's own outputs.
• A benchmark that looked broken was. Belebele scored by log-likelihood put the base model at chance (it always picked "A"). We rescored it with generated letter answers.

Limitations

• Not a source of truth for food safety. It repeats standard guidance (e.g. 74 °C / 165 °F for poultry) but can be wrong; check official sources.
• It does not know historical recipes by heart. It was trained on old cookbooks to read and reason about recipes, not to recall them. For a specific historical dish, give it the text.
• The judges are models. Four open judges from four families agree on the direction and size of the changes, and we keep checking them against people with Label Test, our blind labelling tool. Help refine it: label five answers at label-test.humanjava.com (no account; we keep only a name or initials and your labels).
• Still invents on some no-recipe questions: about 1 in 40 when three of four judges must agree, about 1 in 16 by the strictest single judge. Most remaining cases are borderline hedges (a common default offered with a caveat).
• Strongest in English and Spanish; Thai, Tamil and Hindi had under 200 source records each.
• The old public-domain cookbooks reflect their era: measures, techniques, some outdated advice.

Feedback and security

• The model got something wrong (invented a detail, misread a recipe, unsafe advice): please open a report. Real examples are what improve the next version.
• Help measure it: label five answers on Label Test.
• Security (files that don't match SHA256SUMS, unexpected embedded content): report privately, see SECURITY.md.

Licence and credits

Weights: Apache-2.0, same as the base model. Base model: Qwen3-8B by the Qwen team (Apache-2.0). Training data sources: Wikibooks Cookbook contributors (CC BY-SA 4.0; per-page attribution in the dataset), Project Gutenberg and Internet Archive public-domain books, NHLBI. General chat data: NVIDIA (Nemotron-Instruction-Following-Chat-v1, ODC-BY 1.0, responses by GPT-OSS-120B and Qwen3-235B-A22B), Hugging Face (SmolTalk2, Apache-2.0, responses by Qwen3-32B; prompts from SystemChat-2.0, the Aya Dataset by Cohere For AI, and Table-GPT by Microsoft). Teacher: Qwen3.8-27B (Apache-2.0). Judge: DeepSeek-V4-Flash (MIT). OCR: olmOCR-2 by Ai2 (Apache-2.0). Evaluation: lm-evaluation-harness, XSTest (CC BY 4.0), Belebele (CC BY-SA 4.0). Training: LLaMA-Factory. Quantisation: llama.cpp.

Built by Humanjava Enterprises.