PSompong/hredd-adverse-text-distilbert-onnx

HREDD Adverse-Text Detector — DistilBERT (ONNX student)

A compact, browser-ready text classifier that reads adverse media / allegations and predicts a human-rights or environmental harm category for supplier due-diligence screening. It is the offline "student" reader of the Supplier Risk Screener — distilled from a Llama 3.3 70B teacher and int8-quantized to run entirely in the browser with no API key and no server.

⚠️ Screening aid, not a verdict. This model helps prioritise where human due-diligence attention should go. It does not clear, terminate, or blacklist suppliers, and every High/Critical outcome is meant to be reviewed by a person. It approximates — it does not beat — its 70B teacher (see Evaluation). Built with Llama (see Provenance & licensing).

---

What it does

• Input: a short adverse-text snippet (news headline / allegation, English).
• Output: one of 18 HREDD harm categories — e.g. forced_labour, child_labour, trafficking, wage_theft, osh_nonfatal, fatal_incident, toxic_pollution, deforestation, biodiversity_loss, discrimination, gbvh, retaliation, none.
• The risk tier (Low / Medium / High / Critical) is derived downstream by a severity rubric in the screener — the model emits a harm category, not a tier.

Usage (transformers.js — its primary surface)

import { pipeline } from "@huggingface/transformers";

const detect = await pipeline(
  "text-classification",
  "PSompong/hredd-adverse-text-distilbert-onnx"
);

const out = await detect(
  "Auditors documented withheld wages and confiscated passports at the plant."
);
console.log(out); // e.g. [{ label: "forced_labour", score: 0.91 }]

The repo ships an int8-quantized ONNX graph at onnx/model_quantized.onnx for in-browser inference.

Model details

Field	Value
Base model	distilbert-base-uncased (~66M params, English)
Teacher	Groq Llama 3.3 70B Instruct (llama-3.3-70b-versatile), 5-seed modal labels
Format	ONNX, int8-quantized — ~65 MB (vs ~268 MB full-precision)
Task	Text classification → 18 HREDD harm categories (16 covered by the training corpus)
Tier derivation	harm_category → shared severity rubric → tier (outside this model)
Random seed	42
Training	Free Colab T4 GPU; inverse-frequency class weights; macro-F1 model selection

Training data

Distilled on a 961-row, three-source corpus: synthetic HREDD news snippets, 180 targeted synthetic rows (20 per category, deterministic templates, seed 42), and 54 de-identified curated real cases (each carrying a generic public descriptor). The 30-case evaluation gold set is held out of training. Real company names are not used as labels; synthetic rows carry a [SYNTHETIC NEWS — for portfolio demonstration] marker.

Evaluation (offline, held-out 30-case gold set)

Detection = predicted tier ≥ the gold case's minimum expected tier. Measured fully offline (no API, no key).

Stratum	Bare baseline	This model (offline)	70B teacher
Labour detection (n=23)	1/23 (4.3%)	16/23 (69.6%)	22/23 (95.7%)
Environmental detection (n=7)	—	5/7 (71.4%)	7/7 (100%)
Severe-trio → Critical recall (n=18)	0/18	11/18 (61.1%)	17/18 (94.4%)
Full gold (n=30)	—	21/30 (70.0%)	—

The student recovers roughly two-thirds of the distance from the bare model to the 70B teacher. Its remaining misses are one-tier-short adjacent-family confusions (e.g. forced_labour read as discrimination → still High), not missed harms — and every such case still triggers mandatory human review.

Limitations

• Approximates, does not beat, the teacher — it trades accuracy for offline reproducibility.
• English only (distilbert-base-uncased).
• Single-teacher-family bias — all labels come from one model family (Llama 3.3); its systematic errors transfer to the student.
• Template-overfit risk from the synthetic training rows.
• Out of scope: supplier termination, worker-level scoring, production HREDD without human review, and out-of-domain (financial / medical / legal) text.

Provenance & licensing

• Weights: a fine-tuned DistilBERT student (distilbert-base-uncased, Apache-2.0). This derivative is released under Apache-2.0.
• Built with Llama. The training labels were generated by Meta Llama 3.3 70B (via Groq). Use of those labels is subject to the Llama 3.3 Community License; this model is therefore "Built with Llama."
• Synthetic training rows are marked as demonstration data; curated real cases are de-identified.
• This is a portfolio / methodology artefact — not legal advice or a certification of any supplier.

Links

• Live demo: https://supplier-risk-model.pages.dev
• Part of the Supplier Risk Model — a precautionary HREDD screening tool.