213 lines
7.2 KiB
Markdown
213 lines
7.2 KiB
Markdown
---
|
||
library_name: transformers
|
||
license: apache-2.0
|
||
language:
|
||
- en
|
||
pipeline_tag: text-generation
|
||
base_model: google/gemma-2-2b-it
|
||
tags:
|
||
- gemma
|
||
- gemma-2
|
||
- gdpr
|
||
- compliance
|
||
- legal
|
||
- dpo
|
||
- qlora
|
||
- sft
|
||
datasets:
|
||
- sims2k/GDPR_QA_instruct_dataset
|
||
model-index:
|
||
- name: gdpr_gemma-2-2b
|
||
results:
|
||
- task:
|
||
type: text-generation
|
||
name: GDPR Q&A
|
||
dataset:
|
||
type: sims2k/GDPR_QA_instruct_dataset
|
||
name: GDPR_QA_instruct_dataset
|
||
split: train[:100]
|
||
metrics:
|
||
- type: rouge
|
||
name: ROUGE-L
|
||
value: 0.2252
|
||
- type: bleu
|
||
name: BLEU
|
||
value: 0.1034
|
||
- type: bertscore
|
||
name: BertScore F1
|
||
value: 0.8527
|
||
---
|
||
|
||
# GDPR-Gemma-2-2B — GDPR Compliance Assistant
|
||
|
||
A specialized fine-tune of **`google/gemma-2-2b-it`** for English GDPR
|
||
(General Data Protection Regulation) Q&A. The model is aligned with expert
|
||
GDPR answers via a **3-stage pipeline** — Supervised Fine-Tuning, Dynamic
|
||
Rejection sampling, and Direct Preference Optimization (DPO) — using QLoRA
|
||
for resource-friendly training.
|
||
|
||
> **Disclaimer**: This model provides informational guidance only and **does
|
||
> not constitute legal advice**. Always consult a qualified legal
|
||
> professional for binding GDPR compliance decisions.
|
||
|
||
- 🔗 GitHub: <https://github.com/seok-hee97/gdpr-gemma2>
|
||
- 🧑💻 Author: **seok-hee97** (HF: `cycloevan`)
|
||
- 🏷️ Base: `google/gemma-2-2b-it`
|
||
- 🌐 Language: English
|
||
|
||
---
|
||
|
||
## Training Pipeline (3-Stage)
|
||
|
||
```
|
||
┌──────────────┐ ┌────────────────────┐ ┌──────────────┐
|
||
Base Gemma-2 ─►│ Stage 1: SFT │ ──► │ Stage 2: Dynamic │ ──► │ Stage 3: DPO │
|
||
│ (knowledge) │ │ Rejection Sampling │ │ (alignment) │
|
||
└──────────────┘ └────────────────────┘ └──────────────┘
|
||
```
|
||
|
||
| Stage | Goal | Method |
|
||
|---|---|---|
|
||
| 1. SFT | Inject GDPR domain knowledge | QLoRA SFT on expert Q&A |
|
||
| 2. Dynamic Rejection | Build *realistic* preference pairs | Sample SFT outputs (T=0.9) as `rejected`; expert answer = `chosen` |
|
||
| 3. DPO | Align preferences toward expert answers | DPO on top of SFT adapter (β=0.1) |
|
||
|
||
This pipeline is more faithful than naive DPO because Stage 2 produces
|
||
rejection candidates that match the model's *actual* failure modes, rather
|
||
than synthetic or generic wrong answers.
|
||
|
||
---
|
||
|
||
## Training Configuration
|
||
|
||
| Component | Value |
|
||
|---|---|
|
||
| Base model | `google/gemma-2-2b-it` |
|
||
| Quantization | 4-bit NF4 (QLoRA), bf16 compute |
|
||
| LoRA `r` / `alpha` / `dropout` | 16 / 32 / 0.05 |
|
||
| LoRA target modules | `q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj` |
|
||
| SFT epochs / LR | 3 / 2e-5 |
|
||
| DPO epochs / LR / β | 3 / 5e-6 / 0.1 |
|
||
| Batch size / Grad accum | 1 / 4 |
|
||
| Max prompt / total length | 1024 / 2048 |
|
||
| Optimizer | `paged_adamw_8bit` |
|
||
| Hardware | NVIDIA DGX Spark (CUDA, bf16) |
|
||
|
||
---
|
||
|
||
## Evaluation
|
||
|
||
Quantitative on 100 samples from `sims2k/GDPR_QA_instruct_dataset`;
|
||
qualitative via GPT-4o LLM-as-a-Judge on 10 samples (1–5 scale).
|
||
|
||
### Quantitative (ROUGE / BLEU / BertScore)
|
||
|
||
| Metric | Base | SFT | **DPO (this model)** |
|
||
|---------------|--------|------------|----------------------|
|
||
| ROUGE-L | 0.2072 | **0.2331** | 0.2252 |
|
||
| BLEU | 0.0838 | **0.1146** | 0.1034 |
|
||
| BertScore F1 | 0.8432 | **0.8541** | 0.8527 |
|
||
|
||
### Qualitative (GPT-4o Judge, 1–5)
|
||
|
||
| Criterion | Base | SFT | **DPO (this model)** |
|
||
|-----------------------|------|------|----------------------|
|
||
| Legal Correctness | 3.10 | 3.00 | **3.40** |
|
||
| Article Accuracy | 2.20 | 2.30 | **2.60** |
|
||
| Compliance Alignment | 3.70 | 3.40 | **3.80** |
|
||
| Clarity | **4.10** | **4.10** | 3.80 |
|
||
|
||
DPO improves legal correctness, GDPR-article citation accuracy, and
|
||
compliance alignment over both Base and SFT. It trades a small amount of
|
||
surface-level lexical overlap (ROUGE/BLEU) and clarity in exchange for
|
||
substantively more accurate legal content — a typical alignment trade-off.
|
||
|
||
---
|
||
|
||
## Quickstart
|
||
|
||
```python
|
||
import torch
|
||
from transformers import AutoTokenizer, AutoModelForCausalLM
|
||
|
||
model_id = "cycloevan/gdpr_gemma-2-2b"
|
||
|
||
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
||
model = AutoModelForCausalLM.from_pretrained(
|
||
model_id,
|
||
torch_dtype=torch.bfloat16,
|
||
device_map="auto",
|
||
attn_implementation="eager", # recommended for Gemma-2
|
||
)
|
||
|
||
SYSTEM = (
|
||
"You are a professional GDPR compliance assistant. "
|
||
"Provide accurate, legal, and clear guidance based on the General Data "
|
||
"Protection Regulation."
|
||
)
|
||
|
||
def ask_gdpr(question: str, max_new_tokens: int = 512) -> str:
|
||
messages = [{"role": "user", "content": f"{SYSTEM}\n\nQuestion: {question}"}]
|
||
prompt = tokenizer.apply_chat_template(
|
||
messages, tokenize=False, add_generation_prompt=True
|
||
)
|
||
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
|
||
outputs = model.generate(
|
||
**inputs,
|
||
max_new_tokens=max_new_tokens,
|
||
do_sample=True,
|
||
temperature=0.1,
|
||
top_p=0.2,
|
||
pad_token_id=tokenizer.eos_token_id,
|
||
)
|
||
text = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
||
return text.split("model")[-1].strip() if "model" in text else text
|
||
|
||
print(ask_gdpr("What are the main principles of GDPR?"))
|
||
```
|
||
|
||
---
|
||
|
||
## Intended Use
|
||
|
||
- **In-scope**: Educational explanations of GDPR articles and principles,
|
||
drafting first-pass compliance summaries, internal training material,
|
||
GDPR-aware chatbot prototypes.
|
||
- **Out-of-scope**: Binding legal opinions, jurisdiction-specific advice
|
||
outside the EU/EEA, regulated decisions affecting individuals' rights,
|
||
enforcement/litigation strategy.
|
||
|
||
## Limitations & Risks
|
||
|
||
- **Snapshot of the regulation**: Trained on a static GDPR Q&A dataset;
|
||
does not reflect post-training case law (CJEU rulings, EDPB guidelines)
|
||
or national supervisory authority decisions.
|
||
- **English only**: No multilingual coverage; legal language outside English
|
||
may degrade significantly.
|
||
- **Article-citation accuracy**: Average ~2.6/5 — the model occasionally
|
||
cites incorrect or non-existent article numbers. Always verify citations
|
||
against the official GDPR text.
|
||
- **Alignment trade-off**: DPO improves substantive legal accuracy at a
|
||
small cost to surface fluency vs the SFT-only variant.
|
||
- **Hallucination**: As with any LLM, it can fabricate plausible-looking
|
||
legal references. Treat outputs as drafts, not authoritative sources.
|
||
|
||
## Ethical Considerations
|
||
|
||
GDPR compliance affects individuals' fundamental rights to privacy and data
|
||
protection. Errors in legal interpretation may cause organisations to
|
||
mishandle personal data or mislead data subjects. Use only as a
|
||
decision-support tool, never as the sole basis for compliance actions.
|
||
|
||
## Citation
|
||
|
||
```bibtex
|
||
@misc{gdpr_gemma_2_2b_2024,
|
||
title = {GDPR-Gemma-2-2B: A 3-Stage Aligned GDPR Compliance Assistant},
|
||
author = {seok-hee97},
|
||
year = {2024},
|
||
howpublished = {Hugging Face Model Hub},
|
||
url = {https://huggingface.co/cycloevan/gdpr_gemma-2-2b}
|
||
}
|
||
```
|