Create README.md

README.md

@@ -0,0 +1,121 @@
+---
+license: mit
+datasets:
+- DanielDDDS/recipe-modifications-v2
+language:
+- he
+metrics:
+- f1
+- precision
+- recall
+base_model:
+- dicta-il/dictabert
+pipeline_tag: token-classification
+tags:
+- NER
+- Hebrew
+- recipe
+- CRF
+- DictaBERT
+---
+================================================================================
+README — DanielDDDs/hebrew-recipe-modification-ner
+Model Repository
+https://huggingface.co/DanielDDDs/hebrew-recipe-modification-ner
+================================================================================
+
+OVERVIEW
+--------
+A fine-tuned DictaBERT + CRF model for Named Entity Recognition of recipe
+modifications in Hebrew YouTube cooking comments. The model identifies spans
+where commenters describe substitutions, quantity changes, technique changes,
+and additions to recipes. This checkpoint (P10) is the best-performing
+configuration in a progressive training series evaluated against both human-
+annotated gold examples and silver-labeled data.
+
+--------------------------------------------------------------------------------
+FILE MANIFEST
+--------------------------------------------------------------------------------
+
+best_model.pt                    DictaBERT + CRF model weights.
+                                 Progressive training config: P10.
+                                 Gold F1  : 47.35%  (P 43.94%, R 51.33%)
+                                 Silver F1: 56.05%  (P 56.52%, R 55.58%)
+
+id2label.json                    Integer ID → label string mapping.
+                                 Keys: 0 … 4  →  O, I-SUBSTITUTION,
+                                 I-QUANTITY, I-TECHNIQUE, I-ADDITION
+
+label2id.json                    Label string → integer ID mapping
+                                 (reverse of id2label.json).
+
+training_summary.json            Final training run metrics and
+                                 hyperparameters for the P10 configuration.
+
+evaluation/
+  gold_results.json              Evaluation on the 496-example human gold set.
+                                   F1 : 47.35%
+                                   P  : 43.94%
+                                   R  : 51.33%
+
+  silver_results.json            Evaluation on the silver-labeled test set.
+                                   F1 : 56.05%
+                                   P  : 56.52%
+                                   R  : 55.58%
+
+--------------------------------------------------------------------------------
+MODEL ARCHITECTURE
+--------------------------------------------------------------------------------
+  Base encoder  : DictaBERT (Hebrew BERT trained by the Dicta Institute)
+  Decoder       : Conditional Random Field (CRF) layer
+  Tagging scheme: IO (no B- prefix; spans are contiguous I- sequences)
+  Training data : processed/train_merged.jsonl from the companion dataset repo
+                  (thread-aware tokenization, merged silver + guided splits)
+
+--------------------------------------------------------------------------------
+LABEL SCHEMA
+--------------------------------------------------------------------------------
+  O                 Not a recipe modification span
+  I-SUBSTITUTION    Ingredient or component substitution
+  I-QUANTITY        Quantity or measurement change
+  I-TECHNIQUE       Cooking technique change
+  I-ADDITION        Addition of a new ingredient or step
+
+--------------------------------------------------------------------------------
+PERFORMANCE SUMMARY
+--------------------------------------------------------------------------------
+
+  Evaluation set       Precision   Recall   F1
+  -------------------  ---------   ------   ------
+  Gold  (496 examples)  43.94 %    51.33 %  47.35 %
+  Silver (test split)   56.52 %    55.58 %  56.05 %
+
+  Baseline reference: teacher model upper-bound metrics are available in
+  the companion dataset repo at evaluation/teacher_upper_bound.json.
+
+--------------------------------------------------------------------------------
+USAGE NOTES
+--------------------------------------------------------------------------------
+  • Load best_model.pt with a DictaBERT + CRF inference wrapper.
+  • Use id2label.json / label2id.json to map model outputs to span types.
+  • Input text should be tokenized consistently with the DictaBERT tokenizer
+    used during training (see training_summary.json for tokenizer details).
+  • The model was developed for naturally occurring Hebrew cooking discourse;
+    performance on formal recipe text may differ.
+
+--------------------------------------------------------------------------------
+COMPANION DATASET
+--------------------------------------------------------------------------------
+  DanielDDDs/recipe-modifications-v2
+  https://huggingface.co/datasets/DanielDDDs/recipe-modifications-v2
+
+  Contains raw comment threads, silver labels, gold annotations, full
+  processed splits, and all ablation / P-series training summaries.
+
+--------------------------------------------------------------------------------
+CITATION / CONTACT
+--------------------------------------------------------------------------------
+  Repository owner : DanielDDDs
+  Hugging Face URL : https://huggingface.co/DanielDDDs/hebrew-recipe-modification-ner
+
+================================================================================