A small LoRA adapter that fine-tunes Qwen/Qwen3-8B for vegan and vegetarian recipe Q&A. It was built as the fine-tune stage of an end-to-end LLM pipeline run on workstation hardware, where the real subject was the storage behavior at each stage, not the recipes.

Status: Draft. Companion to the Vegetarian & Vegan Recipe Q&A dataset and the methodology article What I learned building an LLM pipeline on a workstation.

The honest summary first

This is a demo artifact from a storage-infrastructure experiment, not a production recipe tool. A frontier model on a phone writes a better curry than this adapter does, and that was never the point.

What it actually learned: the training set's house style, plus a partial improvement on dietary-constraint adherence. It did not learn to out-cook anything.

So pick by need. If you want a good vegan recipe, use a bigger model. If you want a small, openly-licensed adapter to study or to build on, this is that.

What this is

A LoRA adapter (~87 MB) on top of Qwen/Qwen3-8B. You serve it as base-plus-adapter; it is not a standalone model.
Trained by teacher-student distillation: a 32B teacher (Qwen/Qwen3-32B, run locally) wrote the instruction set, and this 8B student learned from it.
Domain: recipe Q&A only. Tuned for recipe-shaped requests, mediocre at anything else.

Intended use and out of scope

Intended: recipe-style instruction following (give me a recipe, suggest a substitution, adapt a dish to a dietary constraint), and as a small, reproducible adapter for people studying distillation or LoRA serving.

Out of scope: any use where a dietary mistake matters (allergy management, medical, commercial menus) without a second-pass safety filter. See Limitations. Not a general-purpose assistant.

How to use it

The adapter is served on top of the base model with vLLM's LoRA support. The published adapter has had its lm_head LoRA pair stripped so vLLM accepts it (see Notes), so it loads as-is:

bash

vllm serve Qwen/Qwen3-8B \
  --enable-lora \
  --lora-modules vegetarian=knachiketa004/vegetarian-recipe-qwen3-8b-lora \
  --max-lora-rank 16

Then send chat requests with "model": "vegetarian". Two practical notes:

Disable thinking for clean recipe output. Qwen3 defaults thinking on. Pass chat_template_kwargs: {"enable_thinking": false} in the request (or append /no_think to the user message). The adapter was trained with a custom thinking-stripped Qwen3 template, so disabling thinking at serve time matches training behavior.
Do not merge the adapter. On the NeMo Automodel build used here, merge_and_unload() produced an untrained-scoring model on Qwen3 (NeMo Automodel issue #1226, since fixed by PR #1395). Serve adapter-on-base instead, which is the configuration above. For a merged single-model deploy, use a build that includes the #1395 fix and validate the merged outputs against adapter-on-base first.

Training

Setting	Value
Base model	`Qwen/Qwen3-8B` (bf16)
Method	LoRA, all linear layers (q/k/v/o/gate/up/down projections)
LoRA rank / alpha	16 / 32
Global batch size	32
Epochs	2 (~650 steps)
Optimizer / LR	Adam, 1e-4 cosine to 1e-5 min
Sequence length	2048, answer-only loss masking
Tokenizer	Qwen3 chat template, thinking stripped (custom template)
Framework	NeMo Automodel 26.02, FSDP2
Hardware	2x NVIDIA DGX Spark (Grace-class, UMA), single and two-node arms
Training data	train split (10,424 pairs) of the companion dataset (11,582 pairs total)

The fine-tune was one arm of a checkpoint-storage A/B; the storage findings live in the methodology article and its reproduce kit.

Evaluation and limitations

This adapter has real, measured limits. Read them before using it.

Dietary correctness is improved but not solved. On a deliberately hard prompt ("a vegan Punjabi curry with chickpeas and spinach, no coconut") across five sampled seeds at temperature 0.7, the adapter offered coconut oil despite "no coconut" in 2 of 5 samples, versus 4 of 5 for the un-fine-tuned base. Better, not fixed. Hand-verified on a small sample; illustrative, not a benchmark.
It invents inaccurate dish names. On the same prompt it stamped a "Gobi" (cauliflower) label on a chickpea dish in 4 of 5 samples; the base never did. It picked up the training set's habit of naming each dish without the accuracy to make the name correct.
Single-domain and single-language. Recipe Q&A in English only. A recipe-tuned model and a weak general one.
Residual dietary-vocabulary errors. Like the dataset it learned from, it can still slip an animal product into a "vegan" answer by name. For any use where correctness matters, run a second-pass dietary filter (an LLM judge, not a keyword screen, since a keyword filter flags "replace the fish with tofu" as a violation).
Cuisine attribution is imprecise. Regional labels are sometimes wrong even when the dish itself is described correctly.

These bounds inherit from the dataset's documented limitations; the dataset card is the fuller account.

Notes

The lm_head strip. The training recipe applied LoRA to every linear layer, which on Qwen3 includes lm_head. vLLM rejects an adapter whose target_modules include lm_head. The published adapter has that one LoRA pair removed from adapter_model.safetensors and adapter_config.json, which restores clean serving with no meaningful quality loss: the body LoRA across the q/k/v/o/gate/up/down projections carries the adaptation, the output head simply reverts to base, and there was no vocabulary resize.

License and attribution

The adapter weights are released under Apache-2.0, as a derivative of the Apache-2.0 base model (Qwen/Qwen3-8B).

The adapter was trained on the Vegetarian & Vegan Recipe Q&A dataset, which is licensed CC BY-SA 4.0 (it inherits ShareAlike from Wikibooks Cookbook and Wikipedia sources used as grounding and few-shot context). The Apache-2.0 license on these weights does not relicense that dataset: if you redistribute the dataset, its CC BY-SA 4.0 terms apply to it, and attribution to its sources is required. The data sources, via the dataset:

Wikibooks Cookbook (CC BY-SA 3.0), Project Gutenberg (public domain), Wikipedia food articles (CC BY-SA 4.0). The Wikibooks Cookbook material is CC BY-SA 3.0; it is included in the CC BY-SA 4.0 dataset under the one-way compatibility that CC BY-SA 4.0 permits for 3.0 content, so the dataset's 4.0 label and the 3.0 source label are both correct.
Teacher model Qwen/Qwen3-32B (Apache-2.0)

Disclaimer

This is a personal open-source project, provided as-is under the Apache-2.0 license, with no warranty of any kind, express or implied, and no guarantee of fitness for any purpose. You use it at your own risk. As the Limitations section documents, this adapter makes dietary mistakes; do not rely on it for allergy, medical, nutritional, or any other safety-critical decision without independent verification. The author accepts no liability for any use of this model or its outputs. The warranty disclaimer and limitation of liability in the Apache-2.0 license govern; this paragraph restates them in plain language and does not change them.

Citation

bibtex

@misc{nachiketa2026vegetarianrecipeqwen3lora,
  author = {Nachiketa, Kumar},
  title  = {Vegetarian and Vegan Recipe Q&A: a Qwen3-8B LoRA adapter},
  year   = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/knachiketa004/vegetarian-recipe-qwen3-8b-lora}}
}

Reproducibility and companion materials

Methodology article: What I learned building an LLM pipeline on a workstation. Covers the full pipeline (data prep, distillation, fine-tune, eval, serve) and the storage findings at each stage.
Dataset: knachiketa004/vegan-vegetarian-recipes-qa.
Reproduce kit: the as-run training and serving scripts (LoRA recipe, serve command, the lm_head strip) in the AIHomeLab repository.

Acknowledgments

Built on NVIDIA DGX Spark hardware in personal capacity, not affiliated with any employer. On top of vLLM (Apache-2.0), Hugging Face Transformers and PEFT (Apache-2.0), NeMo Automodel, and the Qwen3 model family from Alibaba (Apache-2.0).

This model card is a living document. Issues and suggestions welcome via the AIHomeLab repository.