Ailiance-fr/mascarade-freecad-dataset

Mascarade — FreeCAD / OpenSCAD / CAD parametric Q&A

Description

Q&A bilingue (FR/EN) sur la CAO paramétrique : scripting FreeCAD Python (Part, PartDesign, Sketcher, Draft), code OpenSCAD, CadQuery, modélisation 3D pour impression, design-for-manufacturing.

Ce dataset fait partie de la famille Mascarade, un corpus thématique destiné au fine-tuning LoRA de modèles compacts (cibles : Qwen 2.5-32B, Gemma 3n-E4B, Qwen3 4B) pour des assistants spécialisés en électronique embarquée et CAO mécanique.

Format : JSONL « ShareGPT-style » avec un tour system (rôle d'expert), un tour human (question) et un tour gpt (réponse complète, code compilable, schémas ASCII si pertinent).

Sample count : 4006 examples.

Data sources (EU AI Act Template — AI Office, July 2025)

Publicly available datasets

• —`redcathode/thingiverse-openscad` (CC-BY-SA-3.0) — Thingiverse OpenSCAD models (~2k)
• —`ThomasTheMaker/OpenSCAD` (Various open licenses) — Curated OpenSCAD corpus (~2k)

Web scraping

None directly. All third-party sources are pre-curated HuggingFace datasets above.

Synthetically generated

• —Hand-crafted seed examples (8 rows) authored by the ailiance team using the expert-system-prompt protocol. Marked with metadata.source = "seed" when present.

Licensed data

None.

Sample format

{
  "conversations": [
    {"from": "system", "value": "<expert role prompt>"},
    {"from": "human", "value": "<question>"},
    {"from": "gpt", "value": "<answer with code blocks>"}
  ]
}

Build provenance

Generated by datasets/builders/build_freecad_dataset.py --with-hf --max-samples 2000 from `ailiance/ailiance-models-tuning`. Build date: 2026-05-11. Deterministic — re-running the builder against the same HuggingFace dataset snapshots yields the same output (up to HF source version).

Intended use

• —Fine-tuning small LMs (≤32 B parameters) for freecad domain expertise.
• —Sister datasets in the Mascarade family:
• —Ailiance-fr/mascarade-stm32-dataset
• —Ailiance-fr/mascarade-spice-dataset
• —Ailiance-fr/mascarade-kicad-dataset
• —Ailiance-fr/mascarade-emc-dataset
• —Ailiance-fr/mascarade-embedded-dataset
• —Ailiance-fr/mascarade-iot-dataset
• —Ailiance-fr/mascarade-power-dataset
• —Ailiance-fr/mascarade-dsp-dataset
• —(Mirrored from electron-rare/mascarade-*-dataset.)

License

CC-BY-SA-4.0 — propagated from the most-restrictive third-party source. Adapters trained on this dataset MUST attribute the upstream sources and keep their downstream artifacts under a compatible share-alike license.

PII filter pass — 2026-05-11

A regex-based PII scan was applied to the original 4006 rows on 2026-05-11 (RGPD diligence). Patterns checked: email addresses, AWS/GCP/OpenAI API keys, SSH/PGP private key headers, phone numbers (E.164 + FR), credit card-shaped digit strings, public IPv4. 122 rows containing hard PII (real-looking emails, real phone numbers) were filtered out, leaving 3884 rows (97.0% of original). Soft patterns (GitHub @mentions, Doxygen @brief tags, math constants matching credit-card shapes) were kept as they are not RGPD-sensitive.

The pre-filter version is no longer hosted. Scanner source: tools/pii_scan.py in ailiance/ailiance-models-tuning.

Used to train models evaluated in ailiance/ailiance-bench v0.2

This dataset contributes to training data for hardware-domain LoRA adapters benchmarked in the Ailiance bench suite.

Phase 6 scoreboard verdicts (7-task KiCad/SPICE evaluation):

• —🥇 eu-kiki: champion 4/7 tasks (DSL/PCB/SPICE/extract)
• —🥇 mascarade-embedded: champion P3 extraction (+48 pts)
• —⚠️ mascarade-kicad: catastrophic forgetting on SPICE/P2/P3

See full scoreboard: ailiance-bench README#scoreboard-lora-phase-6.