Instructions to use ronke21/hebnli-dictalm-3.0-24b-lora with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use ronke21/hebnli-dictalm-3.0-24b-lora with PEFT:
from peft import PeftModel from transformers import AutoModelForCausalLM base_model = AutoModelForCausalLM.from_pretrained("dicta-il/DictaLM-3.0-24B-Thinking") model = PeftModel.from_pretrained(base_model, "ronke21/hebnli-dictalm-3.0-24b-lora") - Notebooks
- Google Colab
- Kaggle
DictaLM-3.0-24B (Thinking) HebNLI LoRA adapter
LoRA adapter (rank 16, alpha 32, all linear projection matrices, bfloat16) that adapts dicta-il/DictaLM-3.0-24B-Thinking to 3-class Hebrew NLI on HebNLI. It is one of the Strategy 2 denoisers of the HeRE paper. The model is prompted as
משפט 1: {premise} משפט 2: {hypothesis}
and the label is read from the first generated token (היסק = entailment, סתירה = contradiction, ניטרלי = neutral); in HeRE the soft score is derived from the first-token log-probabilities of the three label tokens.
Training
Learning rate 5e-5, batch size 16, up to 8,000 steps; the checkpoint with the best validation macro-F1 was retained (step 5,280). Single run, seed 42.
Results
| HebNLI dev macro-F1 | HebNLI test macro-F1 | HeRE gold validation F1 | |
|---|---|---|---|
| DictaLM-3.0-24B (Thinking) | 0.876 | 0.902 | 0.861 (Strategy 2, template hypothesis). This adapter produced the finetuned_nli__dictalm24b signal on the full silver corpus and is one of the three models whose unanimous vote (agree3) defines the published HeRE silver label. |
Usage
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
tok = AutoTokenizer.from_pretrained("ronke21/hebnli-dictalm-3.0-24b-lora")
model = AutoModelForCausalLM.from_pretrained("dicta-il/DictaLM-3.0-24B-Thinking", torch_dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(model, "ronke21/hebnli-dictalm-3.0-24b-lora")
License
The adapter is released under the base model's license (apache-2.0); using it requires accepting the base model's terms. Paper resources: dataset https://huggingface.co/datasets/ronke21/HeRE, code https://github.com/Ronke21/HeRE.
Citation
Keinan, R., Cohen, A. D. N., and Tsarfaty, R. (2026). HeRE: A Novel Benchmark for Hebrew Relation Extraction via LLM-Guided Denoising of Knowledge Graph Alignments. In Proceedings of AACL 2026.
- Downloads last month
- 5
Model tree for ronke21/hebnli-dictalm-3.0-24b-lora
Base model
dicta-il/DictaLM-3.0-24B-Thinking