| --- |
| license: apache-2.0 |
| base_model: convaiinnovations/laya |
| language: |
| - en |
| - es |
| library_name: laya |
| tags: |
| - code-search |
| - reranker |
| - system-one |
| - python |
| --- |
| |
| # s1-code v2 |
|
|
| Earlier version of [s1-code](https://huggingface.co/api-service-sac/s1-code-v3). **For use, prefer v3.** A System One |
| decision model (322M, CPU) that returns the probability that a Python function answers a search in English or Spanish. |
| Fine-tuned from [convaiinnovations/laya](https://huggingface.co/convaiinnovations/laya) (Apache 2.0). |
|
|
| Same input format as v3: the state is the file path, the function name, an empty line and the first 1,500 characters |
| of the source; the question is the noul `This code answers the search: <your search>`. |
|
|
| ## Results (new held-out test, 197 questions, 25 candidates from Qwen3-Embedding) |
|
|
| | System | Top 1 | English | Spanish | |
| |---|---|---|---| |
| | s1-code v2 alone | 157 | 80 % | 79 % | |
| | s1-code v2 + Qwen3-Embedding (fused) | 165 | 83 % | 84 % | |
| | Qwen3-Embedding alone | 152 | 76 % | 78 % | |
|
|
| ## Training |
|
|
| About 81,000 questions with exact English/Spanish parity from 306 public repositories, commit messages (CommitPackFT), |
| issues (SWE-bench, SWE-Gym, SWE-smith), CoSQA and CoSQA+ searches, and filtered private functions; one positive and three |
| granite-mined negatives per question; checkpoint taken after the first of two epochs (the second overfitted). |
|
|
| **Known issue:** the CoSQA+ part of its data had false negatives (generated code that also answers the query). They were |
| removed for v3. |
|
|
| ## Licence and attribution |
|
|
| Apache 2.0. Fine-tuned from Laya by Convai Innovations. Training data includes CoSQA+ (CC-BY-4.0, Gong et al.) and CoSQA |
| (MIT). Private training data is not released. |
|
|
