evalexplorer-classify-adapters
Experimental LoRA adapters for the EvalExplorer document classifier, one subfolder per run. Every adapter here is
on an Apache-2.0 base (Qwen3.5, Gemma 4); each keeps its base model's terms. Adapters worth using
on their own get their own repo (baobabtech/evalexplorer-classify-<model>-sft); variants from GRPO and other
trials live here so they do not each need one. Each subfolder holds adapter_config.json,
adapter_model.safetensors, the tokenizer and training_log.json.
| Adapter | Base model | Mean field score | Exact match | Run |
|---|---|---|---|---|
qwen3.5-2b-grpo-countries |
unsloth/Qwen3.5-2B | 0.847 | 0.246 | report |
qwen3.5-2b-grpo-lr5e6 |
unsloth/Qwen3.5-2B | 0.843 | 0.254 | report |
gemma-4-e2b-grpo-lr5e6 |
unsloth/gemma-4-E2B-it | 0.827 | 0.276 | report |
qwen3.5-2b-grpo |
unsloth/Qwen3.5-2B | 0.822 | 0.179 | report |
gemma-4-e2b-grpo |
unsloth/gemma-4-E2B-it | 0.819 | 0.157 | report |
gliner2.5-base-passage |
fastino/gliner2.5-base-v1 | 0.584 | 0.007 | report |
gliner2.5-small-passage |
fastino/gliner2.5-small-v1 | 0.532 | 0.000 | report |
Scores are on the 134-document test split of
baobabtech/evalexplorer-data, config classify_codes. Context and every other
run: baobabtech/evalexplorer-classify-experiments.
from huggingface_hub import snapshot_download
from peft import PeftModel
path = snapshot_download("baobabtech/evalexplorer-classify-adapters", allow_patterns=["gemma-4-e2b-grpo-lr5e6/*"])
model = PeftModel.from_pretrained(base_model, f"{path}/gemma-4-e2b-grpo-lr5e6")
New runs can write here with --output-repo baobabtech/evalexplorer-classify-adapters/<name>.