Retrained on synthetic Apache-2.0 gold; history squashed

4babb9a 27 days ago

2.37 kB

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

eye-grep tagger — electra-small (distilled)

The small, browser-friendly 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.

A google/electra-small-discriminator student (WordPiece), knowledge-distilled from the opsbr/eye-grep-deberta-v3-small teacher (the two have different tokenizers, so distillation is done on aligned per-content-token logits). At ~14 MB int8 it is ideal for an in-browser WebGPU/WASM build — the eye-grep web demo and eye-grep/react use it by default. For maximum accuracy (CLI / server) use the deberta teacher.

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, for WebGPU) and onnx/model_quantized.onnx (int8, for WASM) — plus the WordPiece tokenizer.

Usage

eye-grep React component:

import { LogView } from 'eye-grep/react';
<LogView text={logs} model="opsbr/eye-grep-electra-small" device="webgpu" />

transformers.js (browser):

import { AutoTokenizer, AutoModelForTokenClassification } from '@huggingface/transformers';
const tokenizer = await AutoTokenizer.from_pretrained('opsbr/eye-grep-electra-small');
const model = await AutoModelForTokenClassification.from_pretrained(
  'opsbr/eye-grep-electra-small', { dtype: 'q8', device: 'webgpu' });

Training data

Distilled from the deberta teacher on the synthetic, fully-owned opsbr/eye-grep gold set (Apache-2.0) — no third-party log data.

Notes

A distilled student: smaller and faster than the deberta teacher at some accuracy cost.
Same 11-tag schema and tokenizer-alignment contract as the teacher.