---
license: apache-2.0
base_model: Qwen/Qwen3-8B
library_name: transformers
pipeline_tag: text-generation
language:
- en
tags:
- pii
- redaction
- privacy
- de-identification
- sft
---

# redax-8b

Qwen3-8B fine-tuned to find personally identifying information in text and
return the exact spans. Built as the LLM strategy of redax, a schema-driven
de-identification engine.

## Model Details

- **Developed by:** Dylan Murzello
- **Model type:** causal LM, full-parameter SFT for schema-conditioned span extraction
- **Language:** English
- **License:** Apache-2.0
- **Finetuned from:** [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B)

| file | what |
|---|---|
| `model.safetensors` | bf16 reference weights |
| `redax-8b-Q4_K_M.gguf` | 5 GB, runs on a laptop |
| `redax-8b-Q8_0.gguf` | 8.7 GB, near-lossless |

## Uses

The system prompt names a schema: the labels to find and guard rules for
lookalikes that must be left alone. The model reads the input text and
returns a JSON array of `{"text": ..., "label": ...}` objects — substrings
copied character-for-character, nothing rewritten. When nothing qualifies
it answers `[]`, and it means it: roughly a fifth of the training data is
traps (clinical values, order numbers, codes that look sensitive and are
not).

**Out-of-scope:** this is not a compliance tool. It will miss spans
sometimes, and de-identification regulations (HIPAA, GDPR) are standards a
model cannot certify on its own — keep a human, or at least an ensemble
with pattern matching, in the loop for anything real. English only. Not
for re-identification of individuals.

## How to Get Started

```
ollama pull huggingface.co/dylanmurzello/redax-8b:Q4_K_M
```

One quirk: output opens with an empty `<think></think>` block (Qwen3
training-template artifact). Strip it, then parse the JSON.

## Training Details

53,141 schema-conditioned examples (clinical / financial / general PII),
mixed from public corpora (Nemotron-PII, Gretel) plus targeted synthetic
generation, deduped and 8-gram-decontaminated against the eval benchmark.

| | |
|---|---|
| method | full-parameter SFT (TRL 1.9, assistant-only loss) |
| epochs | 2 (824 steps, packed 2048 ctx, effective batch 32) |
| lr | 1e-5, cosine |
| precision | bf16 |
| final eval loss | 0.0185, no train/eval gap |

## Evaluation

Benchmark rows (strict/relaxed span F1, hard-negative false positives,
privacy leak rate) get added here once the eval suite has run — including
the Q4_K_M vs Q8_0 quantization delta.

## Environmental Impact

One evening on a single rented H100 (~2.5 GPU-hours). The whole fine-tune
cost about as much as a burrito.