owl-int4: code clone detection bi-encoder (INT4)

Offline-first INT4-quantised ModernBERT bi-encoder for code duplicate detection. Runs entirely locally through engram-judge: no Python, no network, no daemon.

Model details

Property	Value
Architecture	ModernBERT-base bi-encoder, 12 layers, 768 hidden, CLS pooling
Quantisation	INT4 group-64, two signed nibbles per byte, per-group scales
Disk	85.4 MB
BigCloneBench F1	0.9592 (F32 baseline 0.9618, delta -0.0026)
Embedding cosine to F32	0.9569
p50 latency	40 ms short pair, 133 ms at BigCloneBench length (Apple Metal)
Peak RSS	~1.6 GB (Metal unified memory, weights dequantised at load)

Usage

export EG_JUDGE_MODEL=/path/to/owl-int4
eg-judge similarity --text-a "$A" --text-b "$B"

{"decision":"duplicate","similarity":0.868,"threshold":0.5,"latency_ms":72.3,"device":"metal"}

eg-judge embed --text "$A" returns the pooled, L2-normalised vector so an index can cache one side and compare with a dot product.

Files

File	Purpose
model.safetensors	INT4 group-64 weights: *.q4 (packed nibbles) plus *.scale per group
encoder/config.json	ModernBERT hyper-parameters
tokenizer/tokenizer.json	BPE tokenizer, CLS pooling
lazy.json	Runtime manifest: weight prefix, pooling mode, head layers

Load with eg-judge similarity --model <this directory>; the loader dequantises at load.

Training data and provenance

Fine-tuned by the checkpoint author on BigCloneBench from the public sentence-transformer Shuu12121/CodeCloneDetection-ModernBERT-Owl. This repository only quantised it; no additional training was performed.

The upstream checkpoint is Apache-2.0, so this derivative carries the same licence and its attribution below.

@misc{shuu12121_codernnclone_modernbert_owl,
  title  = {CodeCloneDetection-ModernBERT-Owl},
  author = {Shuu12121},
  note   = {Sentence-Transformers checkpoint fine-tuned on BigCloneBench},
  year   = {2025}
}

Limits

• Near-miss pairs, which differ by one operator, score F1 0.7139. The pooled vector erases that difference; use a token or AST diff for that band.
• BigCloneBench is Java. Performance on Python, Rust and TypeScript is unmeasured.
• INT4 reduces the file, not the arithmetic: inference runs at the dequantised F16 width.