nimishph commited on
Commit
c6f4050
·
verified ·
1 Parent(s): e4231df

model card

Browse files
Files changed (1) hide show
  1. README.md +29 -0
README.md ADDED
@@ -0,0 +1,29 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: mit
3
+ base_model: jhu-clsp/ettin-encoder-1b
4
+ library_name: onnx
5
+ pipeline_tag: feature-extraction
6
+ tags: [code-search, retrieval, onnx, anvesa]
7
+ ---
8
+
9
+ # anvesa-code-1b
10
+
11
+ A code-search encoder for [anvesa](https://github.com/nimishph/anvesa): Ettin-encoder-1b (jhu-clsp, MIT) fine-tuned
12
+ on 105 MIT-licensed repositories. ONNX, fp32: `model.onnx` with its weights in `model.onnx_data`, and
13
+ `tokenizer.json`. Mean pooling over real tokens, L2-normalised, 512 tokens, no prefix. It embeds anvesa's symbol
14
+ cards without their WQL path line.
15
+
16
+ `ranker-noword.json` is the XGBoost ranker fit on this encoder's scores for anvesa's `experimental.learnedRanker`
17
+ (20 features, no word overlap).
18
+
19
+ | | corpus test (1,710 generated queries, 57 repos, 13 languages) | bench (241 hand-written queries, 8 repos) |
20
+ |---|---|---|
21
+ | first stage, hub penalty 0.3 | 0.352 | 0.685 |
22
+ | with the ranker | 0.370 | 0.699 |
23
+
24
+ CPU: 373 ms a card (ONNX Runtime, 12 threads). For comparison, all-MiniLM-L6-v2 scores 0.293 and 0.547 without a
25
+ ranker.
26
+
27
+ ```bash
28
+ anvesa model install anvesa-code-1b --download
29
+ ```