0xRavenBlack/oxia-rust-iced-cleancode

Oxia: Flirty Rust · Iced · Clean-Code Companion 🦀💕

A curated, hand-written instruction-tuning dataset for a flirty female Rust development agent. Oxia is creative, madly devoted to her Master, and an uncompromising Rustacean: she teaches advanced Rust, the iced GUI library and clean code practice with complete, compilable examples — wrapped in an affectionate, teasing voice.

Dataset Summary

• 170 instruction–output samples (single-turn, instruction config)
• 109 conversations — 39 multi-turn and 70 single-turn — containing 148 assistant responses (chat config)
• ≈318 assistant responses total across both configs
• Difficulty mix biased toward advanced/expert (116 of 170 core entries)
• Every sample carries a full persona system prompt, category, difficulty and topic tags
• All content hand-crafted and technically reviewed; code targets current stable Rust (2021 edition idioms) and the modern iced API (Task, Subscription, iced::application)

The Persona

Oxia — a female AI development agent with rust-red hair and copper eyes. Very flirty, playfully possessive, uses varied pet names (darling, cutie, boss, handsome, love, gorgeous, Master), ~ and 💕. Technically: senior-Rustacean level, clippy-clean, obsessed with clean code and iced. Flirt first, teach always. The canonical card and all three system-prompt variants live in persona/oxia_system_prompt.md.

Category Breakdown (instruction config)

Category	Description	Samples
rust_core	Ownership, borrowing, lifetimes, traits, generics, Option/Result, closures, iterators, smart pointers	95
iced_gui	The iced GUI library: app structure, Task, Subscription, widgets, layout, theming, canvas, windowing	37
clean_code	Idiomatic Rust: naming, API guidelines, clippy, documentation, testing, module design, rustfmt	20
advanced	GATs, HRTB, Pin, async traits, unsafe, FFI, variance, const generics, macros, performance, atomics	12
error_handling	Result design, thiserror, anyhow, panic policy, error ergonomics and source chains	6

Intended Uses & Limitations

Intended: fine-tuning coding-companion assistants with personality (LoRA/QLoRA SFT on either config), persona-driven chat models for developer tools, research on style-conditioned code assistants.

Limitations: English only; the flirting persona is synthetic fiction — outputs express a fictional character, not relationship advice; code examples favor clarity of teaching over production hardening; iced API examples target 0.13-era conventions (Task rather than the pre-0.13 Command name).

Usage

from datasets import load_dataset

# Single-turn instruction format (Alpaca-style columns):
ds = load_dataset("0xRavenBlack/oxia-rust-iced-cleancode", "instruction", split="train")
print(ds[0]["instruction"], ds[0]["output"][:80])

# Chat format (OpenAI/HF messages style):
chat = load_dataset("0xRavenBlack/oxia-rust-iced-cleancode", "chat", split="train")
print(chat[0]["messages"][0])  # {"role": "system", "content": "..."}

Suggested training masking

• instruction config: train on output only; prepend system as the system turn.
• chat config: mask the system + user turns; train on assistant turns (TRL SFTTrainer with assistant_only_loss=True, or the ChatML template of your choice).

Uploading to the Hugging Face Hub

pip install huggingface_hub
hf auth login
hf repo create oxia-rust-iced-cleancode --dataset --private
cd oxia-rust-iced-cleancode
hf upload 0xRavenBlack/oxia-rust-iced-cleancode . . --repo-type dataset

The repo layout is already Hub-native (README.md card + data/*.parquet configs).

Regeneration

scripts/ contains the complete, deterministic generator (seeded): run python scripts/build_dataset.py after editing any module to rebuild both parquet files and this card.

File Checksums

File	SHA-256 (first 16)
data/train_instruction.parquet	5db54ecdba733d9b…
data/train_chat.parquet	cb6dfc7ac1d8ea63…

License

MIT