README.md

---
base_model: mistralai/Mistral-7B-Instruct-v0.2
library_name: peft
pipeline_tag: text-generation
tags:
- lora
- peft
- grc
- compliance
- security
- nist-csf
- cis-controls
license: other
license_name: polyform-noncommercial-1.0.0
license_link: https://polyformproject.org/licenses/noncommercial/1.0.0
---

# ControlSense Adapter

A LoRA adapter that fine-tunes
[mistralai/Mistral-7B-Instruct-v0.2](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2)
to analyze security and governance control statements: it maps a control to
recognized frameworks (NIST CSF 2.0, CIS Controls v8, NIST AI RMF 1.0),
identifies coverage gaps, and emits a structured result suitable for a gap
register.

This adapter is the trained component of **ControlSense**, a locally-running
GRC policy analysis tool. The adapter on its own is roughly 84 MB; it is meant
to be applied to the Mistral base model. For the full application, including
the document ingestion pipeline, UI, and Docker setup, see the project
repository.

- Project: ControlSense (https://benluthy.com)
- Base model: mistralai/Mistral-7B-Instruct-v0.2 (Apache-2.0)

## Intended use

ControlSense is a **decision-support tool**. It is intended to help security
governance practitioners triage and draft control-to-framework mappings and
surface candidate gaps for review. It is not a source of authoritative
compliance determinations, and its output should be reviewed by a qualified
practitioner before being used in any audit, attestation, or compliance
decision.

## How to use

The adapter is designed to be run inside the ControlSense application, which
constructs the prompt format the model was trained on and handles framework
validation of the output. Using the adapter outside that prompt structure will
not reliably produce well-formed results.

To load the adapter directly with PEFT:

```python
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "mistralai/Mistral-7B-Instruct-v0.2"
model = AutoModelForCausalLM.from_pretrained(base, device_map="auto")
model = PeftModel.from_pretrained(model, "MrBenL/controlsense")
tokenizer = AutoTokenizer.from_pretrained("MrBenL/controlsense")
```

## Training

- Method: QLoRA (4-bit) supervised fine-tuning of Mistral-7B-Instruct-v0.2
- Data: a small set of expert-authored GRC examples (approximately 168)
  spanning 12 domains, including access management, data protection, incident
  response, vulnerability management, third-party risk, change management, AI
  governance, LLM security, risk management, security awareness, policies and
  standards, and business continuity
- The training examples are original, authored for this project. They do not
  contain any proprietary, employer, or client material.
- Output is constrained to a structured schema and validated and
  confidence-filtered by the application before display.

## Evaluation

The adapter is evaluated against a held-out set of 100 hand-curated control
examples, each specifying the framework mappings the model should produce and
keywords for the gaps it should surface. On the measured baseline, required-
mapping accuracy is 53.1% (scoped to NIST CSF 2.0, the framework with the
densest training coverage, with each miss expert-adjudicated as a genuine miss
rather than a key error). NIST AI RMF and OWASP LLM mappings are produced and
measured but reported as emerging coverage pending expanded training data. The
product-level fabricated-ID rate is 0.0%: the application's output sanitizer
validates every category against the official framework enumerations and rejects
fabricated or retired IDs before display (the pre-sanitizer model-level rate was
5.2%, almost entirely retired NIST CSF 1.1 numbers). All outputs are intended for
human verification rather than unattended use.

## Limitations

- Coverage is limited to the frameworks and control domains represented in the
  training data; controls outside that scope may be mapped poorly or not at all.
- Like any LLM, the model can produce plausible but incorrect framework
  citations. Confidence scores and source citations are provided to support
  human verification, not to replace it.
- The model is non-deterministic; identical inputs may produce slightly
  different outputs.
- The model can occasionally assign a confident but fabricated code (for example
  a non-existent framework version or an invented subcategory). The output
  sanitizer rejects these structurally, so the affected control is surfaced as
  "No confident mapping" rather than a false mapping, and should be reviewed and
  mapped manually.

## License

This adapter is released under the
[PolyForm Noncommercial License 1.0.0](https://polyformproject.org/licenses/noncommercial/1.0.0).
Use for any noncommercial purpose is free. Commercial use by or on behalf of a
for-profit organization requires a separate commercial license; contact
support@benluthy.com.

The base model, Mistral-7B-Instruct-v0.2, is licensed separately by Mistral AI
under Apache-2.0.

## Citation

This adapter was trained using [TRL](https://github.com/huggingface/trl):

```bibtex
@software{vonwerra2020trl,
  title   = {{TRL: Transformers Reinforcement Learning}},
  author  = {von Werra, Leandro and Belkada, Younes and Tunstall, Lewis and Beeching, Edward and Thrush, Tristan and Lambert, Nathan and Huang, Shengyi and Rasul, Kashif and Gallouédec, Quentin},
  license = {Apache-2.0},
  url     = {https://github.com/huggingface/trl},
  year    = {2020}
}
```