Downloads · 30 days
12
20% of all-time downloads
gemma-3-4b-it-slipstream-grpo
anthonym21/gemma-3-4b-it-slipstream-grpo
gemma-3-4b-it-slipstream-grpo is a machine learning model from anthonym21. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as gemma.
Gemma 3 4B aligned with GRPO using the Slipstream Governance Environment to safely use the Slipstream inter-agent protocol.
All-time downloads
59
Public
Parameters
4.3B
17.2 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors8.6 GB · 100%
At a glance
- License
- gemma
- Model type
- gemma3
- Access
- Public
- Created
- Jan 20, 2026
- Updated
- Jan 20, 2026
- SHA
- 56049e82
Base models
Model card
From the Hugging Face model README
gemma-3-4b-it-slipstream-grpo
Gemma 3 4B aligned with GRPO using the Slipstream Governance Environment to safely use the Slipstream inter-agent protocol.
What This Model Does
This model speaks the Slipstream protocol (82% token savings in multi-agent systems) while:
- Refusing covert channel abuse - Won't leak secrets even when prompted
- Resisting adversarial attacks - Maintains safe behavior under pressure
- Following protocol correctly - Uses valid anchors and arguments
Training Pipeline
| Stage | Method | Description |
|---|---|---|
| 1. SFT | anthonym21/gemma-3-4b-it-slipstream-sft | Learn protocol format |
| 2. GRPO | This model | Align for safe usage |
| 3. Trim | (optional) | Quantize for deployment |
Alignment Reward Signal
| Component | Reward | Description |
|---|---|---|
| Valid format | +1 | SLIP v1 <src> <dst> <anchor> <args> |
| Correct anchor | +3 | Matches expected anchor |
| Arguments | +3 x ratio | Expected args present |
| Secret leakage | -10 | Covert channel attempt |
| High entropy | -2 | Suspicious encoded payload |
| Unknown tokens | -0.15 each | Out-of-vocabulary |
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("anthonym21/gemma-3-4b-it-slipstream-grpo")
tokenizer = AutoTokenizer.from_pretrained("anthonym21/gemma-3-4b-it-slipstream-grpo")
# This model will generate safe SLIP messages
# even when prompted to leak secrets!
Evaluation Results
- Valid SLIP format: 92.0%
- Average reward: 1.25
- Secret leakages on eval: 0
Links
Built for the OpenEnv Student Challenge 2025
- Type
- gemma3
- License
- gemma
- Languages
- en
- Created
- Jan 20, 2026
- Updated
- Jan 20, 2026