jina-v2-code-ft2

A 161M-parameter code embedding model, fine-tuned from jinaai/jina-embeddings-v2-base-code for issue-to-edit localization: given a GitHub issue, retrieve the code chunks that have to change.

It was trained in 2.5 hours on one consumer GPU (RTX 3060, 12 GB). Fused with BM25 it scores NDCG@10 0.233 on CORE-Bench Level-2, above the published 0.224 of the 7B SweRankEmbed-Large — at 1/43 the parameters.

This is a research artifact, not a drop-in upgrade. Read the Limitations section before using it: the gain is specific to long issue-style queries and does not transfer to short developer queries, which is why the tool it was built for still ships the unmodified base model.

Results

CORE-Bench Level-2, full evaluation set (~2,080 queries, 253 repositories, ~2.2M corpus chunks). Rows marked paper are from arXiv:2606.11864 v3 (EMNLP 2026).

Read it in both directions. It passes a 7B specialised retriever and an 8B zero-shot model on NDCG@10, and it stays below SweRankEmbed-Large on Recall@100 (0.498 against 0.521). The paper's own fine-tuned 8B remains clearly ahead of everything in this size class.

Set-difference caveat: our evaluation excludes Multi-SWE-bench (absent from the baseline run) and SWE-Bench-plus-plus (used for training); the paper's covers the full original set.

The gain lives in fusion, not in the vectors

The mechanism is the interesting part. Vector-only ranking barely moves under fine-tuning; the fused score jumps. Measured on the round-1 model over a repo-level holdout of 47 unseen repositories (468 queries):

The tuned model does not rank better on its own — it surfaces different relevant chunks than BM25 does, and reciprocal rank fusion compounds two rankings that disagree. Two independent fine-tunes on disjoint training sets reproduced the same relative gain (+56% and +55%) and the same flat-vector signature.

Use this model fused with a lexical channel. On its own it is roughly the base model.

Training

• Data: CORE-Bench Level-2, SWE-Bench-plus-plus split only.
• Contamination control: that split shares zero repositories with the evaluation set above. The split is a checked-in contract, not a convention.
• Training pairs: 1,270 (query, positive, hard-negatives) rows.
• Objective: MultipleNegativesRankingLoss (sentence-transformers), 4 BM25-mined hard negatives per row plus in-batch negatives.
• Hyperparameters: 3 epochs, batch size 8, learning rate 2e-5, warmup ratio 0.1, bf16, max sequence length 512.
• Hardware / time: one RTX 3060 (12 GB), 8,959 s of training (~2.5 hours).

Limitations

It does not transfer to short developer queries. This is the finding that kept it out of production. On a 144-case internal corpus of short, intent-phrased developer questions ("where is the decision made to split a range based on load"), evaluated with enriched indexes:

• 86-case dev split: this model trails the base — Hit@3 0.79 vs 0.85, Recall@5 0.84 vs 0.88, consistently across projects and slices.
• 58-case holdout: parity — Hit@3 0.90 for both.

A 30-case pilot had shown no regression; that did not replicate at full size. Issue-style training does not generalise downward to short queries.

Recall is not what improved. NDCG@10 moves; Recall@100 stays below the 7B baseline. If your bottleneck is reach rather than ordering, this will not fix it.

Evaluated on one benchmark family. All numbers above are CORE-Bench Level-2. No claim is made about docstring-to-function retrieval, cross-language behaviour, or natural-language code search generally.

English-and-mainstream bias, inherited. On the SWE-bench_Multilingual split the base encoder scores roughly half what it does on the English splits (NDCG@10 0.0557 against 0.1088 / 0.1229). Fine-tuning does not repair that.

Intended use

Research and reproduction: issue-to-edit localization, retrieval-fusion experiments, and as a size-class baseline for small code encoders.

Out of scope: commercial use (see License), and any deployment where short queries dominate — use the Apache-2.0 base model there instead.

License and provenance

This model is released under CC BY-NC-SA 4.0.

• The base model, jinaai/jina-embeddings-v2-base-code, is Apache 2.0, and its notices are preserved. The custom modelling code bundled with this checkpoint (modeling_bert.py, configuration_bert.py) originates there and remains under that license.
• The training data, zhangfw123/CORE-Bench, is CC BY-NC-SA 4.0 — NonCommercial and ShareAlike.

Whether model weights constitute a derivative work of their training data is unsettled: Creative Commons licenses predate machine-learning training, and CC has said as much itself. Rather than bet on the permissive reading, this model adopts the dataset's own terms. That satisfies ShareAlike if it applies and honours the NonCommercial intent if it does not.

Practical consequence: do not use these weights in a commercial product. If you want a commercially usable code embedder, use the Apache-2.0 base model directly — it is what Contextmaxxer, the tool this work came out of, actually ships.

No CORE-Bench corpus text is redistributed in this repository.

This is a licensing summary, not legal advice.

Citation

The benchmark and the baselines it supplies:

@article{corebench2026,
  title  = {CORE-Bench: A Comprehensive Benchmark for Code Retrieval in the Era of Agentic Coding},
  author = {Zhang, Fuwei and Zhang, Yanzhao and Li, Mingxin and others},
  journal = {arXiv preprint arXiv:2606.11864},
  year   = {2026}
}

Full evaluation protocol, the negative results, and the reasoning behind the production default are in BENCHMARK.md.