|
Download README.md from 67L11/CRPR: direct link, hf CLI and curl.
- Browser
- Download file 1.12 kB
-
https://huggingface.co/67L11/CRPR/resolve/main/README.md
- Command line
-
hf download hf://67L11/CRPR/README.md
-
curl -L -o README.md https://huggingface.co/67L11/CRPR/resolve/main/README.md
1.12 kB
| tags: | |
| - calibration | |
| - primaconsis | |
| # PrimaConsis TRC-Cal calibrators | |
| These are small residual calibrators, not full JEV language models. Download the matching upstream backbone separately. Each size folder contains calibrator.pt, pca.pkl and config.json. PCA is mandatory; never mix assets across sizes. | |
| | Size | Source calibration axes | alpha | gamma | delta_max | PCA | | |
| |---|---|---:|---:|---:|---:| | |
| | 2B | D1–D4 | 1.4 | -1.0 | 1.0 | 64 | | |
| | 4B | D1–D3 | 1.4 | -1.0 | 1.0 | 64 | | |
| | 9B | D1–D4 | 1.4 | -1.0 | 1.0 | 64 | | |
| Backbones: alibiserikbay/JevK5-2B, alibiserikbay/JevK5 (4B), alibiserikbay/JevK5-9B. Exact backbone/runtime revisions must be recorded before announcing reproducibility. The source config IDs are references, not a guarantee of compatibility with future upstream main revisions. | |
| Inference uses only the current view's $L-1$ hidden and raw output. Training uses Gold CE plus metric-aligned consistency. The trainer uses only selected source calibration and does not expose target-data injection. | |
| See the GitHub README for install, config, benchmark, train and inference commands. | |