metadata
language: en
license: gemma
base_model: anthonym21/gemma-3-4b-it-slipstream-sft
tags:
- slipstream
- inter-agent-protocol
- grpo
- rlhf
- ai-safety
- gemma-3
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