StephenChou/cedar-raise-qwen3.5-27b

Cedar RAISE Qwen3.5-27B

This repository contains the final RAISE reinforcement-learning LoRA adapter for Cedar access-control policy generation.

• Base model: StephenChou/cedar-qwen27b-sft-v2
• Training run: grpo_v2_27b_oc_m8_b16
• Final checkpoint: step 518 (2 epochs)
• LoRA rank: 32
• LoRA alpha: 64
• LoRA dropout: 0.05

CedarInstruct held-out evaluation

Greedy evaluation on 375 independent scenarios:

Metric	Result
Syntax validity	99.73% (374/375)
Semantic success	48.00% (180/375)
Macro per-check score	81.66%
Micro per-check score	80.83%

Loading

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "StephenChou/cedar-qwen27b-sft-v2"
adapter_id = "StephenChou/cedar-raise-qwen3.5-27b"

tokenizer = AutoTokenizer.from_pretrained(adapter_id)
base = AutoModelForCausalLM.from_pretrained(
    base_id,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(base, adapter_id)
model.eval()

For a standalone model, call model.merge_and_unload() and save the merged model locally.

Intended use

Research on translating natural-language access-control requirements into Cedar policies. Outputs should be validated with the Cedar parser and semantic checks before deployment.

License

Apache-2.0, following the base model.