Hugging Face's logo

Ailiance-fr
/
devstral-python-lora

Text Generation
PEFT
MLX
English
French
lora
ailiance
devstral
python
Model card Files
xet
 Community
devstral-python-lora / README.md
clemsail's picture
clemsail
docs: PST-aligned model card v0.4.2 (EU AI Act Art. 53(1)(d))
969e9f4 verified
preview code
|
raw
history blame
9.69 kB
metadata 
license: apache-2.0
base_model: mistralai/Devstral-Small-2-24B-Instruct-2512
tags:
  - lora
  - peft
  - mlx
  - eu-kiki
  - eu-ai-act
  - art-52
  - art-53
  - gpai-fine-tune
  - pst-aligned
language:
  - en
  - fr
library_name: peft

eu-kiki-devstral-python-lora

LoRA adapter for mistralai/Devstral-Small-2-24B-Instruct-2512, part of the eu-kiki project — a 100 % EU-sovereign multi-model LLM serving pipeline.

EU AI Act compliance posture. This model card is structured to follow the European Commission's Public Summary Template (PST) for the training content of general-purpose AI models, published by the AI Office under Article 53(1)(d) of Regulation (EU) 2024/1689. The structure below (Sections 1–4) maps directly to the PST. Where the official template wording differs from what is reproduced here, the official template wins; please consult the AI Office page for the canonical version. This card is PST-aligned, not PST-verbatim.

Section 1 — General information about the model

Field Value
Model name eu-kiki-devstral-python-lora
Type LoRA adapter (parameter-efficient fine-tune)
Base model mistralai/Devstral-Small-2-24B-Instruct-2512
Provider of the fine-tune L'Électron Rare (Saillant Clément), clemsail
Provider contact https://github.com/L-electron-Rare/eu-kiki/issues
Date of first public release 2026-05-06
Latest version date 2026-05-06
Modalities Text in / text out (no image, audio, or video)
Languages of intended use English, French
Risk classification (EU AI Act) Limited risk (Art. 52)
Systemic-risk class (Art. 51 / 55) Not applicable — this is a LoRA fine-tune, not a foundation model > 10²⁵ FLOPs
Foundation-model provider responsibility The base model provider remains the GPAI provider for the base; this card describes only the fine-tune delta

Section 2 — Description of training content

The following four categories follow the PST four-way classification of training-content sources. Empty categories are listed explicitly so absence is auditable.

2.1 Publicly available datasets

Source URL / Hub ID SPDX licence Records Notes
StarCoder2 Self-Instruct (Python subset) https://huggingface.co/datasets/bigcode/starcoder2-self-align Apache-2.0 2,850 Public HF dataset, Python instruction-tuning pairs

2.2 Data obtained from third parties under licence

No third-party-licensed data used.

2.3 Data collected through web scraping

No web-scraped data used.

2.4 User-provided data and synthetic data

No user-provided or synthetic data used.

Section 3 — Aggregate description of training content

Aggregate field Value
Total records used for this LoRA 2,850
Domain label in the eu-kiki router python
Time-period of source data Mixed; per-source download dates logged in _provenance fields
Modalities in training data Text only
Languages in training data English, French
Estimated total tokens ≈ 570,000 (heuristic 200 tokens / record average)

The full system-level inventory (all 35+ domains across 7 base models / candidates, ≈ 82 K records, with per-source SPDX license, download dates, and n_used counts) is published at docs/eu-ai-act-transparency.md §4.4. This adapter consumes a strict subset of that inventory.

Section 4 — Other relevant elements

4.1 Copyright compliance and TDM opt-out (Art. 53(1)(c))

  • Public datasets (§2.1): all carry permissive open-source licenses (Apache-2.0, MIT, CC-BY-*, BSD); SPDX matrix verified.
  • Third-party-licensed data (§2.2): vendor datasheets used under EU Directive 2019/790 (DSM Directive) Article 4 — Text and Data Mining exception. Robots.txt respected at collection time. SHA-256 manifests published at docs/pdf-compliance-report.md.
  • Scraped data (§2.3): opt-out signals (robots.txt Disallow, <meta name="robots" content="noai">, TDM Reservation headers, ai.txt) are honoured at collection time. Manifests under data/scraped/<source>/manifest.json in the source repo.
  • Removal requests: open an issue at the source repo URL above or contact the operator listed in §1. We commit to remove disputed content within 30 days and re-train the adapter on the next release cycle.

4.2 Quality and curation

  • Per-record _provenance fields (source URL, SPDX license, record_idx, access_date) attached to 49,956 records across 21 domains (system-level), enabling per-record audit and removal.
  • Per-domain cap of ≤ 3 000 records applied to keep classes balanced across the routing surface.
  • Synthetic data (when present) is explicitly marked source: "synthetic" in the row provenance.

4.3 Personal data and PII (Art. 10 + Art. 53(1)(d))

Training data scanned with Microsoft Presidio + en_core_web_lg (2026-04-28) across all 35+ domain directories. One email address detected in the unrelated traduction-tech corpus was redacted before training. No high-signal PII (email, phone, credit card, SSN, IBAN) remains in the released adapters. Low-signal Presidio detections (PERSON, LOCATION, DATE_TIME) are common false positives in technical text and were left in place. Full report: data/pii-scan-report.json in the source repo.

4.4 Special categories of personal data (GDPR Art. 9)

No special-category data (health, religion, sexual orientation, etc.) was intentionally collected. The PII scan above also screens for identifiers that could lead to special-category inference; none were flagged.

4.5 Copyright opt-out registry

The provider tracks opt-outs via the Issues tracker on the source repository. As of release date no removal requests have been received.

Section 5 — Performance evaluation (Art. 53(1)(a))

HumanEval+ (Linux EvalPlus, 164 problems, greedy, 1 sample): base 87.20 / 82.90 → fused +python 86.00 / 81.10. Δ HE+ = −1.80 pts vs base. Scoring on kx6tm-23 (Proxmox PVE 6.17, EvalPlus official sandbox).

Full bench results, methodology, env.json, and rerun.sh per measurement: eval/results/SUMMARY.md · MODEL_CARD.md.

Section 6 — Training configuration

Parameter Value
Method LoRA
Rank 16
Alpha 32
Dropout 0.05
Target modules q_proj, k_proj, v_proj, o_proj (attention only)
Precision BF16
Optimiser AdamW
Learning rate 1e-5
Batch size × grad-accum 1 × 4–8
Framework MLX (mlx_lm fork on Apple Silicon)
Hardware Mac Studio M3 Ultra 512 GB unified memory

6.1 Compute resources (Art. 53(1)(d))

LoRA training is parameter-efficient: only ≈ 0.1–0.5 % of base-model parameters are updated. Estimated training compute ≪ 10²⁵ FLOPs — the systemic-risk threshold of Art. 51. Single-machine training on Mac Studio M3 Ultra; no datacentre footprint. No proprietary teacher model is used in deployed inference.

Section 7 — Usage 

from mlx_lm import load
from mlx_lm.tuner.utils import linear_to_lora_layers
from huggingface_hub import snapshot_download

base_path = snapshot_download("mistralai/Devstral-Small-2-24B-Instruct-2512")
adapter_path = snapshot_download("clemsail/eu-kiki-devstral-python-lora")

model, tokenizer = load(base_path)
linear_to_lora_layers(model, num_layers=32, config={"rank": 16, "alpha": 32})
model.load_weights(f"{adapter_path}/adapters.safetensors", strict=False)

Or fuse and serve as a self-contained checkpoint:

python -m mlx_lm fuse \
    --model mistralai/Devstral-Small-2-24B-Instruct-2512 \
    --adapter-path <adapter_path> \
    --save-path /tmp/eu-kiki-devstral-python-lora-fused \
    --dequantize

Section 8 — Limitations and out-of-scope use

  • Not for safety-critical decisions (medical, legal, structural, life-safety, biometric).
  • Not for high-stakes individual decisions (hiring, credit, law enforcement) — that would re-classify under EU AI Act Art. 6 high-risk and require additional obligations.
  • Hallucination present at typical instruction-tuned LLM levels; pair with a verifier or human-in-the-loop for factual outputs.
  • LoRA inherits all base-model limitations: training cutoff, language coverage, refusal patterns.

Section 9 — Citation 

@misc{eu-kiki-2026,
  title  = {eu-kiki: EU-sovereign multi-model LLM serving with HF-traceable LoRA adapters},
  author = {Saillant, Clément},
  year   = {2026},
  url    = {https://github.com/L-electron-Rare/eu-kiki},
  note   = {Live demo: https://ml.saillant.cc}
}

Section 10 — Changelog

Date Card version Change
2026-05-06 v0.4.1 First HF release — Apache-2.0, EU AI Act self-contained model card
2026-05-06 v0.4.2 Restructured to align with Commission Public Summary Template (PST) §1–4; explicit empty-category disclosure; opt-out registry section added