ten-percent-challenge — Compressed Swiss Multilingual PII Detection Model

Official submission repository by rohitc1612 for the Ten Percent Challenge (DeusExperiments/ten-percent-challenge).

• Hugging Face Repository: rohitc1612/ten-percent-challenge
• Base Model: joelbarmettler/gheim-ch-560m (Swiss Multilingual RoBERTa PII Token Classifier)
• Original FP16 Model Size: 1,118.00 MB
• Packaged Repository Footprint: 109 MB (~103.29 MB raw weights, 9.24% of original)
• Fidelity Retention: 97.35% across 28,872 locked test documents (86,031 gold PII items).
• Feedback Channel: In-model RBF activation steering (zero weight training, 100% bit-identical weights).

---

1. Repository File Structure

All necessary components are self-contained at the repository root:

ten-percent-challenge/
├── config.json          # Model architecture configuration & label map
├── feedback.py          # Online feedback concept ingestion & state builder
├── model.bin            # rANS entropy-packed GPTQ weights & tiered embeddings (~102.5 MB)
├── predict.py           # Standalone CLI batch prediction harness
├── shipcodec.py         # Pure-Python / NumPy rANS decompressor & unpacking kernels
├── shipload.py          # Zero-dependency model reconstructor
├── shiptok.py           # Subword tokenizer wrapper with <unk> fallback
├── tokenizer.json.gz    # Compressed pruned multilingual vocabulary (~0.72 MB)
└── README.md            # Model card & documentation

---

2. Model Description

gheim-ch-560m is a multilingual Swiss RoBERTa token classification model designed to detect Personally Identifiable Information (PII) across Swiss German, French, Italian, and Romansh documents.

This package compresses the 560M-parameter model into 109 MB while maintaining near-baseline detection capabilities across 8 entity types:

• cat:account_number (IBAN / Swiss account numbers)
• cat:private_date (Dates of birth, personal dates)
• cat:private_email (Personal email addresses)
• cat:private_person (Names of private individuals)
• cat:private_phone (Swiss landline and mobile telephone numbers)
• cat:private_url (Personal URLs)
• cat:secret (Passwords, API keys, credentials)
• cat:private_address (Street addresses, Swiss postal codes)

---

3. Compression Architecture

The model achieves an effective storage rate of ~1.46 bits per parameter through four complementary techniques:

1. Sequential Hessian-Weighted GPTQ:
   • Linear layers (Query, Key, Value, Attention Output, FFN) are quantized to mixed 2-bit and 3-bit representations.
   • Activations are ordered by diagonal Hessian magnitude (act_order=True) with exhaustive grid search for optimal clipping thresholds.
2. Asymmetric Numeral Systems (rANS) Bit-Packing:
   • Packed into variable-rate entropy bitstreams that approach theoretical Shannon entropy limits.
3. Domain-Pruned Vocabulary with Zero-<unk> Fallback:
   • Pruned from 250,002 subwords down to ~58,000 Swiss-relevant tokens.
   • Word embeddings stored in tiered precision (3,000 tokens @ 4-bit, 8,000 @ 3-bit, remainder @ 2-bit).
   • Single-character fallback mapping guarantees rare names and addresses are never mapped to <unk>.
4. Latency-Free Zero-Weight Activation Steering:
   • Implements post-deployment human feedback without gradient descent or weight mutation.
   • Computes a latent target centroid and applies an RBF-gated steering vector $w^*$ via PyTorch forward hooks.

---

4. Evaluation & Retention Benchmarks

Evaluated on the locked final exam consisting of 28,872 documents (86,031 gold items) from Swiss court rulings (entscheidsuche), parliamentary records (curia_vista), synthetic PII, and Romansh held-out sources:

Group	$N_{\text{gold}}$	Baseline Acc	Quantized Acc	Retention	Status
ALL ITEMS	86,031	0.883	0.874	97.35%	Overall
cat:account_number	2,357	0.951	0.949	99.64%	PASS
cat:private_date	29,338	0.821	0.824	98.66%	PASS
cat:private_email	2,561	0.992	0.985	99.17%	PASS
cat:private_person	33,002	0.912	0.899	96.49%	PASS
cat:private_phone	3,426	0.940	0.935	99.50%	PASS
cat:private_url	5,323	0.971	0.967	99.17%	PASS
cat:secret	395	0.965	0.952	98.43%	PASS
cat:private_address	9,629	0.856	0.815	93.47%	Sub-par

---

5. Quickstart & Usage

A. Clone the Repository

git clone https://huggingface.co/rohitc1612/ten-percent-challenge
cd ten-percent-challenge

B. Run Inference (CLI)

# Standard batch inference
python predict.py input.jsonl predictions.jsonl

# Steered inference using human feedback
python predict.py input.jsonl adapted_predictions.jsonl feedback_state.pt

C. Ingest Feedback Corrections (CLI)

When human reviewers annotate mistakes (up to 25 spans), generate an updated steering state:

python feedback.py feedback_annotations.jsonl feedback_state.pt

D. Python API Example

import torch
from shipload import load_model
from shiptok import load_tokenizer

# 1. Load model and tokenizer directly from repository root
model = load_model("config.json", "model.bin")
tokenizer = load_tokenizer("tokenizer.json.gz")

# 2. Tokenize and predict
text = "Der Beschwerdeführer Thomas Meier wurde am 14. Januar 1982 in Bern geboren."
inputs = tokenizer(text, return_tensors="pt")

with torch.no_grad():
    logits = model(**inputs).logits
    pred_ids = torch.argmax(logits, dim=-1)[0].tolist()

labels = [model.config.id2label[i] for i in pred_ids]
print(list(zip(tokenizer.convert_ids_to_tokens(inputs["input_ids"][0]), labels)))

---

6. Feedback Verification Protocol

The feedback channel was rigorously audited under double-blind controls:

• Real Feedback Gain: +11.25% Recall (95% CI: [+0.050, +0.175], $p < 0.001$)
• Shuffled Feedback Control: +0.00% Gain (proved zero label memorization)
• Nearby Retention: 100.0% (zero catastrophic degradation on unrelated court rulings)
• Weight Delta: $\Delta W = 0$ (model.bin is completely unmodified)

---

7. Citation & Attribution

@misc{gheim560m_tenpercent,
  title={ten-percent-challenge: 10.8x Compression and Latent Activation Steering for Swiss Multilingual PII Detection},
  author={Rohit (rohitc1612)},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/rohitc1612/ten-percent-challenge}}
}