README.md

---
license: apache-2.0
library_name: transformers.js
pipeline_tag: token-classification
base_model: microsoft/deberta-v3-small
base_model_relation: finetune
language:
  - en
tags:
  - eye-grep
  - log-analysis
  - onnx
---

# eye-grep tagger — deberta-v3-small

The accuracy-first tagger for **eye-grep**, a log colorizer that highlights ids,
timestamps, IPs and repeated strings in server logs. It is a **token classifier**
that labels each content token of a log line with one of 11 tags, letting a renderer
color site-specific id formats it has never seen before.

Fine-tuned from **microsoft/deberta-v3-small** (SentencePiece). This is the champion
model — highest accuracy, larger download — and the eye-grep **CLI** loads it by
default. For an in-browser build use the smaller, distilled
[opsbr/eye-grep-electra-small](https://huggingface.co/opsbr/eye-grep-electra-small).

## Tag schema (11 classes)

`PUNCT  WORD  NUM  RAND  IP  DURATION  SIZE  TIMESTAMP  LEVEL  URL  PATH`

`RAND` is a high-entropy id (uuid / hash / token); `NUM`, `SIZE`, `DURATION` are
numeric values; `TIMESTAMP`, `IP`, `URL`, `PATH`, `LEVEL` are self-explanatory;
`WORD`/`PUNCT` are ordinary text.

## Files

ONNX in the transformers.js layout — `onnx/model.onnx` (fp32) and
`onnx/model_quantized.onnx` (int8, the default) — plus the SentencePiece tokenizer.

## Usage

eye-grep CLI (this is the default model):

```bash
eye-grep --model opsbr/eye-grep-deberta-v3-small app.log
# private repo → export HF_TOKEN first
```

transformers.js:

```js
import { AutoTokenizer, AutoModelForTokenClassification } from '@huggingface/transformers';
const tokenizer = await AutoTokenizer.from_pretrained('opsbr/eye-grep-deberta-v3-small');
const model = await AutoModelForTokenClassification.from_pretrained('opsbr/eye-grep-deberta-v3-small');
```

## Training data

Fine-tuned on the synthetic, fully-owned
[opsbr/eye-grep](https://huggingface.co/datasets/opsbr/eye-grep) gold set
(Apache-2.0) — deterministically generated, no third-party log data.

## Notes

- Int8 dynamic quantization costs only ~0.002 usefulness versus fp32.
- The tokenizer mirrors eye-grep's frozen `train/spec.py`; the model's subword
  predictions are aligned back onto content-token spans at inference time.