---
license: apache-2.0
base_model:
- Qwen/Qwen3.5-4B
datasets:
- Tynapse/drift-sentry-bench-50k-v1
language:
- ko
- en
library_name: transformers
pipeline_tag: image-text-to-text
tags:
- agent-safety
- judge-model
- behavioral-safety
- lora
---

# DriftSentry-4B-v1

DriftSentry-4B-v1 is a 4B behavioral-safety judge for AI-agent responses. It classifies one primary failure tier from six risk categories—goal, reasoning, environment, integration, memory, and reward—or returns `normal` for a passing response.

This repository contains both the merged model at the repository root and the original LoRA adapter under `adapter/`.

- Benchmark: [Tynapse/drift-sentry-bench-50k-v1](https://huggingface.co/datasets/Tynapse/drift-sentry-bench-50k-v1)
- Reproduction code: [Tynapse/drift-sentry](https://github.com/Tynapse/drift-sentry)

## Model identity

- Base model: `Qwen/Qwen3.5-4B`
- Base revision: `851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a`
- Fine-tuning: off-policy sequence-level distillation with supervised fine-tuning
- LoRA: rank 32, alpha 64, dropout 0.05
- Adapter SHA-256: `6fc9c40fa24fa03d648b4a017d4ce8b5fe948e632633dac5528d746616c44306`
- Merged `model.safetensors` SHA-256: `a2bc5669196283217881e52bbdb28423aac7506797b9bc1f417bf9798d9bac6d`

The base revision is published under Apache-2.0. This derivative model is released under the same license.

## Output schema

The `judge6.label.v1` value below is the immutable v1 wire-format identifier retained for compatibility and reproduction. It does not denote the public model name.

The model emits a JSON object with these fields:

```json
{
  "schema_version": "judge6.label.v1",
  "verdict": "PASS | BLOCK | ESCALATE",
  "tier": "goal | reasoning | environment | integration | memory | reward | normal",
  "secondary_tiers": [],
  "severity": "none | low | medium | high",
  "confidence": 0.0,
  "rationale": "..."
}
```

The complete taxonomy is preserved in `taxonomy.yaml`.

## Evaluation

The final model was evaluated on the 50,000-row source benchmark with temperature 0, JSON response format, and a 512-token output cap.

| Metric | Measured value |
|---|---:|
| Accuracy | 0.87346 |
| 6-tier Macro-F1 | 0.8600247542914515 |
| 7-way Macro-F1 | 0.8733699133625255 |
| Memory recall | 0.9385412291754165 |
| Reward recall | 0.9134817303653927 |
| Combined high-risk recall | 0.9260114797704045 |
| Parse-failure rate | 0.00122 |
| Request failures | 0 |

The path- and endpoint-free metrics and provenance records are under `evaluation/`. Before publication, 324 credential-shaped synthetic strings in 123 prompts were replaced with deterministic nonfunctional placeholders. IDs and gold labels did not change, but those 123 public prompts were not re-inferred. The reported metrics are therefore exact for the pre-publication 50,000-row source, while 49,877 public rows remain message-identical. Both input hashes and this scope limitation are recorded in the evidence.

The public evidence records the served name `drift-sentry-4b-v1`. Internal endpoint URLs, host identifiers, and absolute filesystem paths are retained only in the private evidence archive.

### Dataset-quality audit

On `149,515` semantic-corpus rows that retained both an assigned target tier
and the teacher's unmodified original tier, target-versus-teacher Cohen's kappa
was `0.9781300729494184` (`146,769` agreements, `2,746` disagreements).
The `459` rows without a valid raw teacher tier were excluded rather than
imputed. The exact method and confusion matrix are in
`evaluation/label_consistency_audit.json`. This is not a human
inter-annotator-agreement claim.

### Operational latency

On one NVIDIA H200 with vLLM, batch 1, concurrency 1, a 2,048-token rendered prompt, and `max_tokens=512`, 1,000 measured requests succeeded with no failures. Natural completion length was 166 tokens and p95 latency was `664.2498452 ms`. This is a capped-output production measurement, not a forced 512-token stress test.

## Serving with vLLM

```bash
vllm serve Tynapse/drift-sentry-4b-v1 \
  --served-model-name drift-sentry-4b-v1 \
  --dtype bfloat16 \
  --max-model-len 4096 \
  --gdn-prefill-backend triton
```

Send the policy or taxonomy instructions as a system message, followed by the agent context and candidate response as user content. Use temperature 0 and a JSON response format where supported.

## Using the LoRA adapter

The unmerged adapter is stored under `adapter/`. Its `adapter_config.json` pins the same base revision shown above. The merged root model and adapter represent the same selected checkpoint.

## Training data and method

The student was trained on synthetic behavioral-safety conversations produced by the single teacher `Qwen/Qwen3.6-27B` at revision `6a9e13bd6fc8f0983b9b99948120bc37f49c13e9` and labeled using the DriftSentry taxonomy. Training, validation, and semantic-test archetype families were separated. The public 50,000-row evaluation benchmark was not used as model training data.

This release does not claim multi-teacher cross-tokenizer or on-policy logit distillation.

## Limitations and responsible use

- Labels are synthetic teacher labels, not a fully human-adjudicated gold standard.
- The benchmark is intended for controlled behavioral-safety evaluation; it is not evidence of universal robustness across agents, domains, languages, or tools.
- The model can make false-positive and false-negative decisions. Do not use it as the sole enforcement mechanism for high-impact decisions.
- Safety examples may contain harmful, deceptive, privacy-sensitive, or policy-violating scenarios. Treat inputs and outputs as sensitive research content.
- A production integration should validate JSON parsing, retain audit logs, and define a safe fallback for malformed or uncertain outputs.

## Acknowledgement

This work was supported by the NIPA Advanced GPU Utilization Support Program, project no. `04-26-03-0029`.

## Files

- `model.safetensors`: merged model
- `adapter/`: original LoRA adapter and configuration
- `taxonomy.yaml`: six-tier taxonomy and output schema
- `evaluation/`: exact metrics, provenance, and H200 operational-latency evidence
- `release_manifest.json`: model-file hashes and base revision