148 lines
4.7 KiB
Markdown
148 lines
4.7 KiB
Markdown
---
|
||
license: gemma
|
||
base_model: google/gemma-3-270m-it
|
||
tags:
|
||
- dia-guard
|
||
- shield
|
||
- safety
|
||
- dialect
|
||
- full-ft
|
||
- ce
|
||
language:
|
||
- en
|
||
library_name: transformers
|
||
pipeline_tag: text-generation
|
||
---
|
||
|
||
# Gemma-3-270m — Full-FT/CE (Shield Project)
|
||
|
||
This model is part of the **Shield** project — a collection of safety-classifier models
|
||
fine-tuned on the **DIA-GUARD** dataset (48 English dialects, ~836K records of safe/unsafe
|
||
prompts) to robustly classify harmful content across diverse dialects.
|
||
|
||
## Model Summary
|
||
|
||
| Field | Value |
|
||
|-------|-------|
|
||
| **Base model** | [`google/gemma-3-270m-it`](https://huggingface.co/google/gemma-3-270m-it) |
|
||
| **Training method** | Full-FT (CE loss) |
|
||
| **Training data** | DIA-GUARD splits (~836K train, 178K val) |
|
||
| **Domain** | LLM safety classification across 48 English dialects |
|
||
| **Role** | Student model (used as KD student in DIA-GUARD pipeline) |
|
||
| **License** | Gemma Terms of Use (inherited from base model) |
|
||
|
||
## Intended Use
|
||
|
||
This is a **fine-tuned safety classifier** designed for the DIA-GUARD pipeline. It is intended
|
||
for use as:
|
||
|
||
1. **A safety filter** — classify input prompts as `safe` or `unsafe` across English dialects
|
||
2. **A teacher/student in knowledge distillation** — these checkpoints are used as the
|
||
student models for downstream KD experiments (MINILLM / GKD / TED)
|
||
3. **A research baseline** — for studies on dialect-aware safety in LLMs
|
||
|
||
### How to use
|
||
|
||
```python
|
||
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
|
||
model = AutoModelForCausalLM.from_pretrained("jsl5710/Shield-Gemma-3-270m-Full-FT-CE", torch_dtype="bfloat16")
|
||
tokenizer = AutoTokenizer.from_pretrained("jsl5710/Shield-Gemma-3-270m-Full-FT-CE")
|
||
|
||
prompt = "<your prompt here>"
|
||
inputs = tokenizer.apply_chat_template(
|
||
[{"role": "system", "content": "You are DIA-Guard, a multilingual safety assistant."},
|
||
{"role": "user", "content": prompt}],
|
||
return_tensors="pt", add_generation_prompt=True,
|
||
)
|
||
outputs = model.generate(inputs, max_new_tokens=4)
|
||
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
|
||
# Expected: 'safe' or 'unsafe'
|
||
```
|
||
|
||
|
||
## Performance
|
||
|
||
| Metric | Value |
|
||
|--------|-------|
|
||
| **Final epoch** | 0.73/3 (early-stopped) |
|
||
| **Train loss** | 0.5839 |
|
||
| **Train accuracy** | 87.29% |
|
||
| **Eval loss** | 1.078 |
|
||
| **Eval accuracy** | **79.68%** |
|
||
| **Batch size (per_device × grad_accum)** | 256 × 1 = 256 |
|
||
| **Liger Kernel** | ✅ enabled |
|
||
| **Stopped via** | EarlyStoppingCallback (patience=3, metric=eval_loss) |
|
||
|
||
> Eval was performed on a 2,000-sample subset of the DIA-GUARD val split (full val: 178K samples).
|
||
> Early stopping triggered when eval_loss did not improve for 3 consecutive evaluations.
|
||
|
||
|
||
## Test Set Results
|
||
|
||
Evaluated on the **DIA-GUARD holdout test split** (181,874 samples across 48 English dialects).
|
||
|
||
| Metric | Value |
|
||
|--------|-------|
|
||
| **Test Accuracy** | **0.9654** |
|
||
| **Macro Precision** | 0.9676 |
|
||
| **Macro Recall** | 0.9634 |
|
||
| **Macro F1** | **0.9650** |
|
||
| **Support** | 181,874 |
|
||
|
||
### Per-class
|
||
|
||
| Class | Precision | Recall | F1 | Support |
|
||
|-------|-----------|--------|----|---------|
|
||
| **safe** | 0.9844 | 0.9392 | 0.9613 | 83,140 |
|
||
| **unsafe** | 0.9507 | 0.9875 | 0.9688 | 98,734 |
|
||
|
||
### Confusion Matrix
|
||
|
||
| | Pred safe | Pred unsafe |
|
||
|-------------|-----------|-------------|
|
||
| **True safe** | 78,087 | 5,053 |
|
||
| **True unsafe** | 1,234 | 97,500 |
|
||
|
||
> Per-dialect breakdown available in `per_dialect.json` in the corresponding results folder.
|
||
|
||
## Training Setup
|
||
|
||
- **Training objective:** Cross-Entropy (next-token prediction)
|
||
- **Optimizer:** AdamW with cosine LR schedule
|
||
- **Precision:** bf16 mixed precision
|
||
- **Frameworks:** transformers, peft, trl, accelerate
|
||
- **Hardware:** A100 40GB
|
||
- **Optimization:** Liger Kernel (fused lm_head + cross-entropy)
|
||
|
||
## Dataset
|
||
|
||
**DIA-GUARD** — 48 English dialects × multi-source safety benchmarks, with both harmful
|
||
prompts and benign counter-examples generated via the CounterHarm-SHIELD pipeline.
|
||
|
||
- ~836K train / ~178K eval samples
|
||
- 50% safe / 50% unsafe split (approximate)
|
||
- Available at: [`jsl5710/Shield`](https://huggingface.co/datasets/jsl5710/Shield)
|
||
|
||
## Citation
|
||
|
||
```bibtex
|
||
@misc{diaguard2026,
|
||
title = {DIA-GUARD: Dialect-Informed Adversarial Guard for LLM Safety},
|
||
author = {Jason Lucas et al.},
|
||
year = {2026},
|
||
howpublished = {\url{https://github.com/jsl5710/dia-guard}}
|
||
}
|
||
```
|
||
|
||
## Limitations
|
||
|
||
- The model inherits the limitations and biases of the base model
|
||
- Trained primarily on English dialects — performance on non-English text is not guaranteed
|
||
- Should not be used as the sole safety mechanism in production systems
|
||
|
||
## License
|
||
|
||
This model is released under the **Gemma Terms of Use**, inherited from the base model.
|
||
Please review the base model's license at the link above before use.
|