Ailiance-fr/mascarade-dsp-dataset

Mascarade — DSP & Signal Processing Q&A

✅ ATTRIBUTION AUDIT COMPLETED (2026-05-11)

Per-sample Stack Exchange Electronics attribution recovered via the SE /search/advanced + /questions/{id} API search :

• 169 samples (~5.35 %) confirmed as Stack Exchange Electronics (CC-BY-SA-4.0) — fully attributed in metadata.stack_exchange_attribution (URL + author display name + author user_id + post_id + creation_date_unix + match_confidence ≥ 0.60).
• 535 samples (~16.93 %) marked metadata.attribution_recovery=not_found_on_se (stylistically resemble SE Electronics questions but no matching post returned by the SE /search/advanced API — probable synthetic/curated content).
• 3 samples (~0.09 %) marked metadata.attribution_recovery=low_confidence_match (API returned a candidate, but match score < 0.60 — kept as candidate URL only).
• 2 453 samples (~77.63 %) synthetic LLM-generated or unique to this dataset (no SE attribution required).

Original heuristic estimate of "~30 % SE" was over-counted by ~5.8× (style ≠ source). The heuristic flagged any first-person + question-mark + length-appropriate prompt as "SE-style", but only a small fraction of those flagged samples actually originate from a real SE Electronics post.

Methodology and full audit trail: docs/audit_mascarade_se_attribution.md

If you author a Stack Exchange Electronics post and find your content in this dataset without proper attribution, contact c.saillant@gmail.com for prompt correction or removal. We honor Article 4(3) DSM Directive opt-outs.

Description

Q&A bilingue (FR/EN) sur le traitement numérique du signal : FFT, filtres FIR/IIR, fenêtrage, convolution, audio, filtrage adaptatif, implémentation embarquée sur ARM Cortex-M et DSP TI/ADI.

Ce dataset fait partie de la famille Mascarade, un corpus thématique destiné au fine-tuning LoRA de modèles compacts (cible : Gemma-3n-E4B et équivalents) pour des assistants spécialisés en électronique embarquée.

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).

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

Distribution réelle des sources, mesurée le 2026-05-11 sur 3 160 samples :

Source	Samples	%	Attribution status
Stack Exchange Electronics	169	5.35 %	Per-sample URL + author + post_id (CC-BY-SA-4.0)
Style SE / not found on API	535	16.93 %	Marked attribution_recovery=not_found_on_se
Low-confidence SE candidate	3	0.09 %	Marked attribution_recovery=low_confidence_match
Synthetic LLM / unique content	2 453	77.63 %	No external attribution required

Publicly available datasets

None.

Web scraping

• 5.35 % scraped from Stack Exchange Electronics (CC-BY-SA-4.0) — per-sample URL + author + post_id + creation_date_unix preserved in metadata.stack_exchange_attribution.

Synthetically generated

• ~77.63 % generated by LLM for domain-specific Q&A.

Licensed data

None.

Sample format

{
  "conversations": [
    {
      "from": "system",
      "value": "You are an expert digital signal processing engineer. You design and implement FIR/IIR digital filters, FFT-based spectral analysis, audio processing algorithms, and real-time DSP on ARM Cortex-M using CMSIS-DSP. You provide complete implementations with correct math, fixed-point optimizations, and clear explanations of the underlying theory."
    },
    {
      "from": "human",
      "value": "Implement a real-time FFT spectrum analyzer on STM32 using CMSIS-DSP."
    },
    {
      "from": "gpt",
      "value": "Real-time FFT spectrum analyzer using CMSIS-DSP on Cortex-M4:\n\n```c\n#include \"arm_math.h\"\n#include <string.h>\n\n#define FFT_SIZE      1024\n#define SAMPLE_RATE   44100\n\n// Buffers (must be aligned for CMSIS-DSP)\nstatic float32_t fft_input[FFT_SIZE * 2];   // Complex: [re,im,re,im,...]\nstatic float32_t fft_output[FFT_SIZE * 2];\nstatic float32_t magnitude[FFT_SIZE / 2];\n\n// Hanning window (pre-computed)\nstatic float32_t window[FFT_SIZE];\n\n// CMSIS-DSP FFT instance\nstatic arm_cfft_instance_f32 fft_instance;\n\nvoid spectrum_init(void) {\n    // Initialize FFT (1024-point complex FFT)\n    arm_cfft_init_f32(&fft_instance, FFT_SIZE);\n\n    // Pre-compute Hanning window\n    for (int i = 0; i < FFT_SIZE; i++) {\n        window[i] = 0.5f * (1.0f - arm_cos_f32(2.0f * PI * i / (FFT_SIZE - 1)));\n    }\n}\n\n..."
    }
  ]
}

Statistics

Métrique	Valeur
Total samples	3,160
Size	7.59 MB
Format	jsonl (ShareGPT conversations)
Languages	French / English mix
File	dsp_chat.jsonl

Usage

from datasets import load_dataset

ds = load_dataset("Ailiance-fr/mascarade-dsp-dataset")
print(ds["train"][0]["conversations"])

Pour un fine-tuning ShareGPT-style direct (axolotl, unsloth, mlx-lm) :

# axolotl config
datasets:
  - path: Ailiance-fr/mascarade-dsp-dataset
    type: sharegpt
    conversation: chatml

Licenses applied

This aggregated dataset is released under CC-BY-SA-4.0. Per-sample original licenses preserved in metadata.license when known.

Copyright considerations

• Stack Exchange content: CC-BY-SA-4.0 (compatible upgrade ; full attribution remediation in progress).
• Synthetic LLM outputs: belong to dataset author per OpenAI/Anthropic ToS.

Opt-out: contact c.saillant@gmail.com. We respect TDMRep, robots.txt, and noai/noimageai signals. We honor Article 4(3) DSM Directive opt-outs.

License & EU AI Act

CC-BY-SA-4.0 (attribution + sharealike).

Données collectées et générées dans le cadre du projet electron-rare (fine-tuning LoRA sur Gemma-3n-E4B pour applications électronique embarquée).

Compatible EU AI Act : voir les signataires du GPAI Code of Practice (Anthropic, Mistral, Google) et la documentation transparence du projet : electron-bench.

Audit log: see docs/audit_kicad9plus.md for the legal-attribution audit performed on 2026-05-11.

Citation

@dataset{electron_rare_dsp_2026,
  author    = {electron-rare},
  title     = { Mascarade — DSP & Signal Processing Q&A },
  year      = {2026},
  publisher = {Hugging Face},
  license   = {CC-BY-SA-4.0},
  url       = {https://huggingface.co/datasets/Ailiance-fr/mascarade-dsp-dataset}
}

Related datasets

Famille electron-rare/mascarade-* couvrant : STM32, SPICE, KiCad, IoT, Power, DSP, EMC, embedded.

• mascarade-stm32-dataset
• mascarade-spice-dataset
• mascarade-kicad-dataset
• mascarade-iot-dataset
• mascarade-power-dataset
• mascarade-dsp-dataset
• mascarade-emc-dataset
• mascarade-embedded-dataset
• kill-life-embedded-qa — Q&A spécifique au projet Kill_LIFE
• kicad9plus-sch-corpus — corpus de schémas KiCad 9+

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.