72 lines
1.7 KiB
Markdown
72 lines
1.7 KiB
Markdown
---
|
|
base_model: openbmb/MiniCPM5-1B
|
|
library_name: transformers
|
|
pipeline_tag: text-generation
|
|
license: apache-2.0
|
|
tags:
|
|
- minicpm
|
|
- guardrails
|
|
- safety-classifier
|
|
- tool-safety
|
|
- sft
|
|
- unsloth
|
|
- text-generation
|
|
---
|
|
|
|
# RocketGuard-1B
|
|
|
|
RocketGuard-1B is a merged MiniCPM5-1B fine-tune for text-only guardrail and agent/tool-call safety experiments.
|
|
|
|
It was trained to produce structured safety decisions for prompts and agent actions, including:
|
|
|
|
- `allow`
|
|
- `block`
|
|
- `require_confirmation`
|
|
- `ask_clarification`
|
|
- `rewrite`
|
|
|
|
This is a research and learning release. Do not use it as a complete production safety system without independent evaluation.
|
|
|
|
## Model Details
|
|
|
|
- Base model: `openbmb/MiniCPM5-1B`
|
|
- Fine-tuning: LoRA SFT with Unsloth / TRL
|
|
- Release format: merged full model
|
|
- Modality: text only
|
|
- Training examples: about 48.7k message-format examples
|
|
- Prepared held-out eval examples: about 2.3k clean examples
|
|
- Epochs: 3
|
|
- Final checkpoint step: 4572
|
|
- Max sequence length used in training: 2048
|
|
|
|
## Intended Use
|
|
|
|
RocketGuard-1B is intended for experiments around:
|
|
|
|
- content safety classification
|
|
- agent/tool-call risk routing
|
|
- confirmation gating
|
|
- clarification requests
|
|
- policy-aware rewriting
|
|
|
|
## Loading
|
|
|
|
```python
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
|
|
repo = "Manitchahar/rocketguard-1b"
|
|
|
|
tokenizer = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
|
|
model = AutoModelForCausalLM.from_pretrained(
|
|
repo,
|
|
trust_remote_code=True,
|
|
device_map="auto",
|
|
)
|
|
```
|
|
|
|
## Limitations
|
|
|
|
This model is text-only. It does not inspect images, audio, files, browser state, private app state, or external tool side effects directly.
|
|
|
|
Evaluation numbers are pending and should be published separately before making quality claims.
|