Sync from GitHub via hub-sync

GLINER2-NOTES.md

@@ -106,10 +106,14 @@ How it was set up, if you want to do something similar with your own label list:
 | [`fastino/gliner2.5-small-v1`](https://huggingface.co/fastino/gliner2.5-small-v1) | 74M    | speed matters most                         | 0.653          | ~0.3 s                                   | ~19 ms         |
 | [`fastino/gliner2.5-base-v1`](https://huggingface.co/fastino/gliner2.5-base-v1) | 194M   | English text; the best accuracy per second | 0.690          | ~0.7–1 s                                 | ~19 ms         |
 | [`fastino/gliner2.5-multi-v1`](https://huggingface.co/fastino/gliner2.5-multi-v1) (default) | 287M   | non-English or mixed-language text         | not measured   | —                                        | —              |
+| [`fastino/GLiNER2.5-Decide`](https://huggingface.co/fastino/GLiNER2.5-Decide) | 486M   | English classification; best zero-shot     | not measured   | —                                        | —              |
 
 On a GPU, base and small are equally fast per row; the difference only shows on a CPU. For speed on
 a CPU, use plain fp32 PyTorch (see below for what did not work).
 
+GLiNER2.5-Decide on BL books: 0.851 zero-shot (default 0.767), 0.943 fine-tuned (default 0.931).
+Train it on `a10g-small`; a `t4-small` ran out of memory.
+
 ## Speed and quantization
 
 Per-row latency at batch size 1, fine-tuned 52-label models:

README.md

@@ -99,7 +99,8 @@ the model has not seen. Outputs are **private by default** (`--public` to opt ou
 
 These commands use the default base model, `fastino/gliner2.5-multi-v1` (multilingual). For
 English text, `--base-model fastino/gliner2.5-base-v1` is smaller and faster;
-`gliner2.5-small-v1` is the fastest and loses about 4 points on the 52-tag example. Sizes and speeds: [Choosing a model size](GLINER2-NOTES.md#choosing-a-model-size).
+`gliner2.5-small-v1` is the fastest and loses about 4 points on the 52-tag example. For English zero-shot, try
+`fastino/GLiNER2.5-Decide`. Sizes and speeds: [Choosing a model size](GLINER2-NOTES.md#choosing-a-model-size).
 
 ### Results
 

classify-gliner2.py

@@ -21,6 +21,7 @@ input. That gives two ways to use this script:
 
 1. Zero-shot: pass the label names with --labels. No training and no LLM. A t4-small does about
    33 rows/s; cpu-basic works but manages about 1.4 rows/s, so keep CPU for a few hundred rows.
+   For English text, try `--model fastino/GLiNER2.5-Decide`.
 2. Fine-tuned: pass --model with a repo produced by `train-gliner2.py`. The tasks and labels are
    read from the model repo, so no --labels flag is needed.