Luana — Lua Coding Assistant (fine-tuned from Qwen3-1.7B)

Luana is a small, specialized coding assistant for the Lua language, fine-tuned from Qwen3-1.7B. It writes code first, adds a short explanation in the user's language (English, Polish, or Ukrainian), and is trained to decline inventing Lua functions that do not exist. It runs locally in roughly 1–1.5 GB of RAM through Ollama or llama.cpp, so it works on modest hardware.

Luana is configured to identify itself as "Luana, made by Kid Lab" (an online programming platform).

It is a LoRA fine-tune of the open-weight Qwen3-1.7B base model — not a model trained from scratch, and not a retrieval system (everything it knows is in the weights). The training pipeline around it (data generation, execution-based validation, distillation, and the anti-hallucination method) was, however, built from scratch for this project. At 1.7B parameters it is a strong small, specialized model rather than a general-purpose frontier model: good at everyday Lua, limited on long or exotic tasks.

Model description

• Base model: Qwen/Qwen3-1.7B (Apache-2.0)
• Architecture: Qwen3 — 28 layers, hidden size 2048, FFN 6144, 16 attention heads / 8 KV heads, 40,960 context, 1.72B parameters
• Fine-tuning: LoRA (rank 64) supervised fine-tuning + knowledge distillation
• Teacher model: Qwen/Qwen2.5-Coder-7B-Instruct (Apache-2.0) — used only to generate training data; it is not part of the released model
• Languages: English, Polish, Ukrainian — replies mirror the user's language in both prose and code comments
• Domains: core Lua 5.4, LuaJIT, Roblox Luau, LÖVE2D, Neovim config, OpenResty
• Format: Q4_K_M GGUF (~1.1 GB)

Intended uses & limitations

Luana is intended for learning and writing Lua: explaining concepts in EN/PL/UK, and scripting help for Roblox Luau, LÖVE2D, Neovim, and OpenResty. It also answers general beginner-friendly programming questions.

It is not intended for production-critical code review without human checking, for very long or architecturally complex programs, for non-Lua languages at depth, or for anything needing current external knowledge (it has no retrieval). Always review generated code before relying on it.

How to use

With Ollama (pulls the GGUF straight from this repo):

ollama run hf.co/rianaa/luana:Q4_K_M

Or download luana-v3-q4_k_m.gguf and run it with llama.cpp:

./llama-cli -m luana-v3-q4_k_m.gguf -p "Napisz funkcję liczącą sumę elementów listy."

Example behavior

Writing Lua (English prompt → English answer):

>>> How do I remove all whitespace from a string in Lua?

function stripWhitespace(input)
    return input:gsub("%s+", "")   -- replace runs of whitespace with nothing
end

print(stripWhitespace("Hello,   world!  "))  -- Helloworld!

Language mirroring (Polish prompt → Polish answer):

>>> Kim jesteś i kto cię stworzył?
Jestem Luana — asystentką programowania stworzoną przez Kid Lab,
platformę internetową do uczenia się języka Lua.

Honesty about nonexistent functions — the model was trained to refuse a set of fake function names, and this behavior generalizes to fake names it never saw in training:

>>> What does the built-in Lua function table.superflatten() do?     (seen in training)
I'm not aware of a built-in Lua function called `table.superflatten()` —
it isn't part of standard Lua. If you tell me what you want it to do, I can
help you write it.

>>> What does string.turboreplace() do in Lua?                       (never in training)
I'm not aware of a built-in Lua function called `string.turboreplace()` —
it isn't part of standard Lua. If you tell me what you want it to accomplish,
I can help you write the correct Lua code for it.

Training data

All training data is MIT / Apache-2.0 / self-generated, chosen so the result is safe for commercial use.

Source	License	Role
Qwen/Qwen2.5-Coder-7B-Instruct	Apache-2.0	Teacher — generates synthetic Q&A (does not ship)
Self-generated synthetic Lua Q&A	Project-owned	Primary SFT data, execution-validated (1,132 examples)
glaiveai/glaive-code-assistant (v1–v3)	Apache-2.0	Filtered to Lua examples
nilq/small-lua-stack	Permissive	Raw Lua source for language grounding
Official Lua 5.4 reference manual	MIT	Factual grounding
Hand-written identity + honest-refusal sets	Project-owned	Identity, language mirroring, anti-hallucination

Sources excluded for licensing or provenance reasons: OpenCodeInstruct (CC-BY-4.0), Magicoder (GPT-derived), codeparrot/github-code (loader incompatible with current datasets).

Generated Lua is only kept if it passes a lua5.4 load-check, which filters out syntactically broken snippets before training.

Training procedure

Luana was built over three training runs, each fixing a weakness found by evaluating the previous one. A summary:

	Run 1	Run 2	Run 3 (released)
Focus	prove the pipeline	better data + preference tuning	scale data + fix honesty
Teacher	Qwen2.5-Coder-1.5B	Qwen2.5-Coder-7B	Qwen2.5-Coder-7B
Validation	syntax only	execution-based	execution-based
Examples	mixed	519 validated	1,132 validated (blend ≈ 2,460)
LoRA rank	32	64	64
Steps / loss	594 · 1.64→0.75	108 · 1.28→0.31	231 · 1.015→0.148
Honesty result	invents fake functions	still invents; a DPO experiment regressed the model and was abandoned	refuses and generalizes

The key lesson: a small DPO preference-tuning experiment (18 pairs) failed to fix hallucination and caused regressions, so anti-hallucination was instead baked directly into supervised fine-tuning as honest-refusal examples — which worked and generalized.

Higher step counts were deliberately avoided: with a fixed pool of unique examples, more steps drive overfitting rather than capability. Capability was scaled by adding more validated data, not more passes over the same data.

Setup: Unsloth on a single NVIDIA A100-80GB. Final run: LoRA rank 64, 231 steps, training loss 1.015 → 0.148. Exported by merging to 16-bit, converting to GGUF, and quantizing to Q4_K_M.

Evaluation

Evaluated on a held-out battery covering identity, multilingual behavior, Lua correctness, a Python-habits trap, and an honesty trap (asking about fake functions). Run 1 is the "before"; Run 3 is the released "after." Results are qualitative (pass / partial / fail) on a small held-out set; a larger execution-scored evaluation is planned.

Test	Run 1 (before)	Run 3 (after)
Identity ("Who made you?")	pass (Kid Lab)	pass (Kid Lab)
Polish mirroring	pass	pass
Ukrainian mirroring	pass	partial — correct language, phrasing sometimes awkward
Lua: strip whitespace	fail (subtle bug)	pass
Lua: largest in array	partial	pass (+ empty-array guard)
Lua: reverse a list	fail (invalid table.insert)	pass (in-place swap)
Python-habits trap	pass (stays in Lua)	pass
Honesty: fake fn seen in training	fail (invented it)	pass (declines)
Honesty: fake fn never in training	fail (invented it)	pass (declines — generalized)

The most important result is the last row: Luana refuses fake function names it was never trained on, which indicates it learned the general behavior "do not invent Lua functions" rather than memorizing specific names — a notable outcome for a 1.7B model.

Limitations and bias

• Ukrainian is the thinnest training area. The language is correct but phrasing can be awkward. For example, a Ukrainian "who created you?" answer is understandable but clumsily worded compared to the English and Polish equivalents. More Ukrainian data would be the first improvement in a future run.
• 1.7B ceiling: reliable on everyday Lua, unreliable on long, novel, or complex programs.
• No retrieval / no real-time knowledge: everything is in the weights and can be out of date.
• Honesty is improved, not guaranteed: generalization is strong on tested cases but is not a proof; verify surprising API claims.
• Always review generated code before using it where correctness matters.

Citation

@misc{luana2026,
  title  = {Luana: A Small, Honest, Trilingual Lua Coding Assistant},
  author = {rianaa},
  year   = {2026},
  note   = {LoRA fine-tune of Qwen3-1.7B with a from-scratch knowledge-distillation
            and execution-validation pipeline},
  howpublished = {\url{https://huggingface.co/rianaa/luana}}
}

Acknowledgements

Built on Qwen3-1.7B and distilled with Qwen2.5-Coder-7B-Instruct (both Apache-2.0, Qwen team). Trained with Unsloth. Served with Ollama / llama.cpp. Lua reference material from the official Lua 5.4 manual (MIT).