--- tags: - sentence-transformers - feature-extraction - code-search - code-embedding - retrieval - modernbert - dense base_model: Shuu12121/NightOwl pipeline_tag: feature-extraction library_name: sentence-transformers license: apache-2.0 datasets: - Shuu12121/coir_hard_negative_datasets_v3_kd - Shuu12121/owl_code_search_hard_negative_datasets_V2_kd - Shuu12121/codeedit_hard_negative_datasets_kd --- # NightOwl-CodeEmbedding 🦉 `NightOwl-CodeEmbedding` is a compact, 768-dimensional dense embedding model built for code retrieval, code-edit retrieval, and technical question answering. It is fine-tuned from [`Shuu12121/NightOwl`](https://huggingface.co/Shuu12121/NightOwl), a ModernBERT-based code encoder. The model uses CLS pooling with cosine similarity and does **not** require `query:` / `passage:` style prefixes. ## Highlights * Compact (150.8M parameters) yet competitive on CoIR-style code retrieval benchmarks * Ranks **18th out of 241 models overall** on the [MTEB(Code, v1)](https://huggingface.co/spaces/mteb/leaderboard) leaderboard, and is the **top-scoring dense single-vector model under 300M parameters** among scored entries — ahead of many models an order of magnitude larger (see [Leaderboard Standing](#leaderboard-standing)) * Standard single-vector retrieval: one 768-d vector per document and plain dot-product / cosine search, so it drops straight into any existing vector-database stack * Covers **eight programming languages**: the six CodeSearchNet languages plus Rust and TypeScript * Handles a broad range of retrieval scenarios: NL-to-code search, code-to-code retrieval, **code-edit retrieval**, and technical QA * Trained with hard negatives mined by `Qwen/Qwen3-Embedding-0.6B` (15 hard negatives per anchor) * Decontaminated against the CodeSearchNet test splits and the CodeEditSearchRetrieval benchmark (see [Data Decontamination](#data-decontamination)) * Drop-in compatible with `sentence-transformers`; Apache-2.0 licensed ## Supported Languages * Go, Java, JavaScript, PHP, Python, Ruby (CodeSearchNet languages) * **Rust, TypeScript** (additional) Performance on languages outside this set is untested and may vary. ## Usage ```python from sentence_transformers import SentenceTransformer model = SentenceTransformer("Shuu12121/NightOwl-CodeEmbedding") queries = ["Python function that sorts a list in descending order"] documents = [ "def sort_desc(values): return sorted(values, reverse=True)", "def average(values): return sum(values) / len(values)", ] query_embeddings = model.encode(queries) document_embeddings = model.encode(documents) # Cosine similarity (embeddings are normalized internally by similarity()) scores = model.similarity(query_embeddings, document_embeddings) print(scores) ``` ## Model Details | Property | Value | | ----------------------- | -------------------- | | Base model | `Shuu12121/NightOwl` | | Architecture | ModernBERT | | Parameters | 150,779,136 | | Embedding dimension | 768 | | Pooling | CLS pooling | | Maximum sequence length | 1,024 tokens | | Similarity | Cosine similarity | | Query/document prefixes | Not required | | Weight dtype | FP32 | | Weight memory | 575 MiB | | License | Apache-2.0 | ## MTEB Results The model was evaluated with MTEB on code-related retrieval and technical QA tasks. Evaluation setup: * Model revision: `c7c8a57b9539297e192d5cf39b9aecf1fb376edd` * MTEB version: `2.15.1` * Metric: `NDCG@10` * Hardware: NVIDIA GeForce RTX 5090 * Batch size: 64 Multi-subset task scores are reported as macro averages. | Task | Split | NDCG@10 | | -------------------------------- | ------: | ----------: | | AppsRetrieval | test | 0.39177 | | COIRCodeSearchNetRetrieval | test | 0.84264 | | CodeEditSearchRetrieval | train¹ | 0.74808 | | CodeFeedbackMT | test | 0.76690 | | CodeFeedbackST | test | 0.85207 | | CodeSearchNetCCRetrieval | test | 0.91805 | | CodeSearchNetRetrieval | test | 0.89239 | | CodeTransOceanContest | test | 0.75953 | | CodeTransOceanDL | test | 0.36057 | | CosQA | test | 0.42810 | | StackOverflowQA | test | 0.86608 | | SyntheticText2SQL | test | 0.68266 | | **Macro average, all 12 tasks** | | **0.70907** | | **CoIR macro average, 10 tasks** | | **0.68684** | ¹ `CodeEditSearchRetrieval` does not provide a standard `test` split in MTEB, so the official `train` split is used for evaluation. These examples were **not** used for fine-tuning. See [Data Decontamination](#data-decontamination). ### Leaderboard Standing On the public [MTEB(Code, v1)](https://huggingface.co/spaces/mteb/leaderboard) leaderboard, the model's task-mean score (70.91, i.e. the 12-task macro average above ×100) places it as follows: * **#18 of 241 models overall**, ahead of many models an order of magnitude larger * **#6 of 155 among sub-1B-parameter dense single-vector models** — and the **smallest model in that top six**. The five models ranked above it (`F2LLM-v2-0.6B/330M`, `pplx-embed-v1-0.6b`, `C2LLM-0.5B`, `Qwen3-Embedding-0.6B`) are all roughly 0.33–0.6B parameters, i.e. 2–4× larger. * **#1 among ranked dense single-vector models under 300M parameters** (the leaderboard's small-model view) > **Reading the numbers fairly.** MTEB(Code, v1) reports a *zero-shot %* for each > model — the fraction of leaderboard tasks the model was *not* trained on. > `NightOwl-CodeEmbedding` is **8%** zero-shot: it was trained on most of these task > families, so its score reflects strong **in-domain** retrieval rather than zero-shot > transfer. Models marked **100%** (e.g. `embeddinggemma-300m`, the > `granite-embedding` r2 family, `Qwen3-Embedding`) are evaluated fully out-of-domain, > so raw score comparisons across rows with different zero-shot % are not > apples-to-apples. The fairest direct comparisons are to other code-specialized > models at similar zero-shot levels (e.g. the `F2LLM` / `C2LLM` families at 8–58%). ### Comparison with similar-sized models The table below compares `NightOwl-CodeEmbedding` with other compact **dense single-vector (bi-encoder)** models on MTEB(Code, v1), with a size ladder of larger single-vector models for reference. Score is the leaderboard task mean (higher is better); the *Zero-shot* column is the share of tasks the model did not train on. | Model | Params | Emb. dim | Max tokens | Zero-shot | MTEB(Code, v1) ↑ | | ---------------------------------------------------- | ------: | -------- | ---------: | --------: | ---------------: | | **`NightOwl-CodeEmbedding`** (this model) | 150.8M | 768 | 1,024 | 8% | **70.91** | | `codefuse-ai/F2LLM-v2-160M` | 159M | 640 | 40,960 | 58% | 70.38 | | `google/embeddinggemma-300m` | 308M | 768 | 2,048 | 100% | 68.76 | | `codefuse-ai/F2LLM-v2-80M` | 80M | 320 | 40,960 | 58% | 67.97 | | `ibm-granite/granite-embedding-311m-multilingual-r2` | 312M | 768 | 8,192 | 100% | 63.84 | | _Larger single-vector reference (size ladder)_ | | | | | | | `codefuse-ai/F2LLM-v2-0.6B` (#1 sub-1B) | 596M | 1,024 | 40,960 | 58% | 77.41 | | `Qwen/Qwen3-Embedding-0.6B` | 596M | 1,024 | 32,768 | 100% | 75.42 | | `codefuse-ai/F2LLM-v2-14B` (#1 overall) | 13.99B | 5,120 | 40,960 | 58% | 80.75 | Takeaways: * Among compact single-vector dense models, `NightOwl-CodeEmbedding` is the strongest entry in the leaderboard's small-model view while also being one of the smallest, edging out `F2LLM-v2-160M` and clearly ahead of `embeddinggemma-300m`. * The sub-1B leaders (`F2LLM-v2-0.6B`, `Qwen3-Embedding-0.6B`) score ~4–6.5 points higher but are ~4× the parameter count with larger embedding dimensions, which directly increases index size and inference cost. * The 14B model at the top of the overall board is ~10 points higher but ~93× larger — a different deployment cost regime entirely. ## Base Model: the NightOwl Backbone `NightOwl-CodeEmbedding` is fine-tuned from [`Shuu12121/NightOwl`](https://huggingface.co/Shuu12121/NightOwl), a ModernBERT-style code encoder that was **pre-trained from scratch** — tokenizer included — rather than adapted from a general-purpose checkpoint. The whole stack, from tokenization to the pre-training objective, is designed around code. **Code-aware tokenizer.** NightOwl uses a custom 50,368-token BPE tokenizer in which whitespace is tokenized **independently** of adjacent words, so indentation gets its own tokens instead of being merged into "leading-whitespace + word" pieces. In code, the same identifier recurs at many indentation depths; folding whitespace into those pieces would waste large parts of the vocabulary on near-duplicate "indent + token" variants. Keeping whitespace separate lets the fixed vocabulary budget cover more genuinely distinct subwords while still representing indentation faithfully — which matters for whitespace-significant languages such as Python. **Two-phase pre-training with line-level masking.** NightOwl is trained with masked-language modeling (`mlm_probability = 0.3`) in two phases: * *Phase 1 — mixed pre-training:* standard random-token MLM over code, natural language, and technical documentation (producing `NightOwl-Pre`). * *Phase 2 — code-only continuation:* **line-level MLM**, where entire source-code lines are masked instead of random tokens. This aligns the pre-training objective with code search and retrieval, where the unit of meaning is closer to a line or statement than an isolated token. The recommended `NightOwl` checkpoint is this Phase-2 result. Backbone architecture (base): | Property | Value | | ------------------------------ | ----------------------------------------------------- | | Architecture | ModernBERT (alternating local/global attention, RoPE) | | Parameters | ≈150M | | `hidden_size` / layers / heads | 768 / 19 / 12 | | Vocabulary | 50,368 (custom code BPE) | | Max sequence length | 1,024 (Phase 1) → 2,048 (Phase 2) | Pre-training data mixes `bigcode/starcoder2data-extras` (Kaggle notebooks, StackOverflow threads, GitHub issues, technical documentation, …) with whole-file source from `Shuu12121/github-file-programs-dataset` across the eight supported languages. Long examples are split into chunks so all tokens are used rather than truncated. As a raw backbone — before any embedding fine-tuning — NightOwl reaches **0.8436 average MRR** on MTEB `CodeSearchNetRetrieval` under a fixed SentenceTransformer fine-tuning protocol, ahead of CodeBERT-base (0.7944), GraphCodeBERT-base (0.8325), UniXCoder-base (0.8003), and ModernBERT-base (0.8182) evaluated the same way. `NightOwl-CodeEmbedding` builds the retrieval model described in this card on top of that backbone. ## Training The model was trained with `CachedMultipleNegativesRankingLoss` using bidirectional query-to-document and document-to-query objectives. | Property | Value | | -------------------------- | ------------------------------------ | | Training samples | 2,534,400 | | Positives per anchor | 1 | | Negatives per anchor | 15 | | Loss | `CachedMultipleNegativesRankingLoss` | | Objective | Bidirectional retrieval training | | Hard-negative mining model | `Qwen/Qwen3-Embedding-0.6B` | | Epochs | 1 | | Learning rate | 6e-5 | | Batch size | 1024 | ### Training Data The training data is a mixture of: 1. **Public code-retrieval datasets** covering the following CoIR task families: AppsRetrieval, COIRCodeSearchNetRetrieval, CodeFeedbackMT, CodeFeedbackST, CodeSearchNetCCRetrieval, CodeSearchNetRetrieval, CodeTransOceanContest, CodeTransOceanDL, CosQA, StackOverflowQA, and SyntheticText2SQL. 2. **Custom code–comment pair data**: code snippets paired with natural-language description comments across the eight supported languages. 3. **Code-edit data** derived from `commitpackft`, pairing edit intents with code changes. All datasets were constructed as hard-negative retrieval datasets: for each anchor, one positive and fifteen hard negatives. Hard negatives were mined with [`Qwen/Qwen3-Embedding-0.6B`](https://huggingface.co/Qwen/Qwen3-Embedding-0.6B), which retrieves semantically similar but non-matching candidates — considerably harder than random negatives. The mining model is used only during dataset construction and is not required at inference time. This setup is intended to sharpen discrimination between code snippets, programming questions, edit examples, and technically similar retrieval candidates. ### Data Decontamination To reduce benchmark contamination, the following overlaps were removed from the training data **before** training: * Overlaps between the custom code–comment pair data and the **CodeSearchNet test split** * Overlaps between the `commitpackft`-derived code-edit data and the **CodeEditSearchRetrieval** benchmark evaluation data For `CodeEditSearchRetrieval`, note that MTEB labels the evaluation split `train`. This is just the official split name for the task; the evaluated examples were not included in this model's fine-tuning data. The reported score should therefore be read as **in-domain generalization on held-out benchmark examples** — not training-set performance, though, given the in-domain training distribution, not strictly zero-shot performance either. ## Intended Use * Natural language to code search * Code-to-code retrieval and similar-function search * Code-edit retrieval (matching edit intents to code changes) * Retrieval over programming Q&A and technical questions * Local semantic code search systems * RAG systems over codebases and developer documentation Example use cases include indexing functions, snippets, programming solutions, StackOverflow-style answers, code-review examples, and edit-related code examples. ## Limitations * Specialized for code-related retrieval; may underperform general-purpose text embedding models on unrelated natural-language tasks. * Inputs longer than 1,024 tokens are truncated — a shorter context window than some competitors (e.g. the 8K+ token `F2LLM` and `granite` models), so very long files must be chunked. * MTEB(Code, v1) is largely in-domain for this model (8% zero-shot). On code domains, query styles, or languages far from the training distribution, expect lower performance than the leaderboard numbers suggest. * Performance may vary by programming language, query style, and the granularity of indexed code chunks; languages outside the eight supported ones are untested. * The model produces dense single-vector embeddings. For applications that need very fine-grained token-level matching, other retrieval approaches — such as late-interaction (multi-vector) models or cross-encoder rerankers — may be worth considering, with different trade-offs in index size and retrieval infrastructure. ## Recommended Indexing Settings Encode both queries and documents with normalized embeddings: ```python embeddings = model.encode(texts, normalize_embeddings=True) ``` With normalized embeddings, dot product is equivalent to cosine similarity. For codebase search, indexing function-level or class-level chunks usually works best. Very long files may exceed the 1,024-token context limit and should be split into smaller semantic chunks. ## Citation If you use this model, please cite it together with the base model and Sentence Transformers. ```bibtex @misc{nightowl_codeembedding, title = {NightOwl-CodeEmbedding}, author = {Shuu12121}, year = {2026}, publisher = {Hugging Face}, url = {https://huggingface.co/Shuu12121/NightOwl-CodeEmbedding} } ```