laya-uni-2

laya-uni-2: retrained with about 4,400 real, openly licensed course PDFs (all faculties; German, English, French, Italian) on top of the synthetic data. File Sorter moves to it on its own within a day (with Updates on).

• Real course PDFs (240, 13 faculties, never trained on): right module first 85.0 % (laya-uni-1 78.8 %, base 59.6 %). At p ≥ 0.6 and 2× the next it sorts 81.7 % of them (laya-uni-1 62.5 %) at 93.9 % precision. Mathematics 66.7 % (was 46.7 %).
• Unseen degree programs (855 synthetic files): right module first 84.1 % (was 81.8 %).
• Thresholds in the manifest: p ≥ 0.75 and 2× the next (95 % precision on validation).

See the README for training, evaluation, limits and credits (Apache-2.0, derived from convaiinnovations/laya).

file	SHA-256
model.safetensors	2b48bd423d1ed84f1745f73f8964588997f771756f1e75481197d46681df9f90
tokenizer.json	609d8f4c067cd3950f88594c5a802616cea245823836ef5848ee4fc40aab5b6f
rl_agent_config.json	15c84ac6f5339ddbb6e7cfed4f996e70116bfecbf9e3e4344865eec139c93a66
encoder-config.json	36b7d4898274ed6f06ce9c1f0e34034d8bd39295ab232d7cabc41a3eac821e5f

laya-uni 1

First release of laya-uni: Laya (multilingual) fine-tuned to pick the university module of a downloaded file, for every faculty, in German, English, French and Italian.

• Unseen degree programs: right module first 81.8 % (base 57.5 %); 93.2 % precision on what it sorts at p ≥ 0.6 / 2× lead.
• Real course PDFs (240, 13 faculties): right module first 78.8 % (base 59.6 %); 94.7 % precision.

See the README for training, evaluation, limits and credits (Apache-2.0, derived from convaiinnovations/laya).

file	SHA-256
model.safetensors	2a73a52757ff1557d59abcdb43db469327df077146ce68140292a943572925bb
tokenizer.json	609d8f4c067cd3950f88594c5a802616cea245823836ef5848ee4fc40aab5b6f
rl_agent_config.json	c06f9dff24ed9f1132e5e44020381d380d3b5ea30e650f90a82ff6378215c6ff
encoder-config.json	36b7d4898274ed6f06ce9c1f0e34034d8bd39295ab232d7cabc41a3eac821e5f