model card

README.md

@@ -0,0 +1,29 @@
+---
+license: mit
+base_model: jhu-clsp/ettin-encoder-150m
+library_name: onnx
+pipeline_tag: feature-extraction
+tags: [code-search, retrieval, onnx, anvesa]
+---
+
+# anvesa-code-150m
+
+A code-search encoder for [anvesa](https://github.com/nimishph/anvesa): Ettin-encoder-150m (jhu-clsp, MIT) fine-tuned
+on 105 MIT-licensed repositories, learning from the 1B fine-tune as a teacher. ONNX, fp32: `model.onnx` with its weights in `model.onnx_data`, and
+`tokenizer.json`. Mean pooling over real tokens, L2-normalised, 512 tokens, no prefix. It embeds anvesa's symbol
+cards without their WQL path line.
+
+`ranker-noword.json` is the XGBoost ranker fit on this encoder's scores for anvesa's `experimental.learnedRanker`
+(20 features, no word overlap).
+
+| | corpus test (1,710 generated queries, 57 repos, 13 languages) | bench (241 hand-written queries, 8 repos) |
+|---|---|---|
+| first stage, hub penalty 0.3 | 0.301 | 0.626 |
+| with the ranker | 0.322 | 0.632 |
+
+CPU: 40 ms a card (ONNX Runtime, 12 threads). For comparison, all-MiniLM-L6-v2 scores 0.293 and 0.547 without a
+ranker.
+
+```bash
+anvesa model install anvesa-code-150m --download
+```