Add DeBERTa Large ANLI model

Upload the fine tuned DeBERTa Large checkpoint, tokenizer, evaluation results, and model documentation.

README.md

@@ -1,3 +1,167 @@
 ---
 license: cc-by-nc-4.0
+language:
+- en
+base_model: microsoft/deberta-large
+datasets:
+- facebook/anli
+pipeline_tag: text-classification
+tags:
+- deberta
+- anli
+- natural-language-inference
+- sequence-classification
+- transformers
+metrics:
+- accuracy
+model-index:
+- name: DeBERTa Large ANLI
+  results:
+  - task:
+      type: text-classification
+      name: Natural Language Inference
+    dataset:
+      type: facebook/anli
+      name: ANLI combined rounds
+      split: dev
+    metrics:
+    - type: accuracy
+      value: 0.6221875
+      name: Validation accuracy
+  - task:
+      type: text-classification
+      name: Natural Language Inference
+    dataset:
+      type: facebook/anli
+      name: ANLI combined rounds
+      split: test
+    metrics:
+    - type: accuracy
+      value: 0.6178125
+      name: Test accuracy
 ---
+
+# DeBERTa Large fine tuned on ANLI
+
+## Model
+
+This checkpoint is based on `microsoft/deberta-large`.
+
+It was fine tuned only on the combined ANLI training rounds. The training set contained 162,865 premise and hypothesis pairs.
+
+The model predicts one of three labels:
+
+| Label | Meaning |
+|---:|---|
+| 0 | entailment |
+| 1 | neutral |
+| 2 | contradiction |
+
+The input order is premise first and hypothesis second.
+
+## Evaluation
+
+Accuracy was measured on the combined ANLI held out rounds.
+
+| Split | Accuracy | Examples |
+|---|---:|---:|
+| Development | 62.22% | 3,200 |
+| Test | 61.78% | 3,200 |
+
+These values are plain classification accuracy.
+
+The checkpoint was trained on ANLI alone. Comparisons should use the same combined ANLI splits and the same label mapping. The results are not presented here as a universal leaderboard claim.
+
+The machine readable results are stored in `baseline_eval.json`.
+
+## Training
+
+| Setting | Value |
+|---|---:|
+| Base model | `microsoft/deberta-large` |
+| Epochs | 3 |
+| Batch size | 16 |
+| Gradient accumulation steps | 2 |
+| Learning rate | 0.000016958369168519958 |
+| Weight decay | 0.1 |
+| Warmup ratio | 0.1828387398995507 |
+| Label smoothing | 0.1 |
+| Maximum sequence length | 128 |
+| Seed | 1299843651 |
+| Numerical precision | FP32 |
+
+The hyperparameters were selected for this model and dataset combination.
+
+The full training record is stored in `model_card.json`.
+
+## Use
+
+Load the repository with `AutoTokenizer` and `AutoModelForSequenceClassification` from the Transformers library.
+
+Pass the premise and hypothesis as a text pair.
+
+Use a maximum sequence length of 128 to match training.
+
+## Files
+
+| File | Purpose |
+|---|---|
+| `model.safetensors` | Model weights |
+| `config.json` | Architecture and label mapping |
+| `tokenizer.json` | Tokenizer data |
+| `tokenizer_config.json` | Tokenizer settings |
+| `baseline_eval.json` | Evaluation results |
+| `model_card.json` | Training record and provenance |
+| `README.md` | Model card |
+
+## Limitations
+
+The model was trained and evaluated on English ANLI data.
+
+ANLI is adversarial and difficult. Performance on other NLI datasets may differ.
+
+The training accuracy was 98.57%, while held out accuracy was lower. This gap should be considered when using the checkpoint.
+
+The model can inherit errors and biases from the base model and the training data.
+
+The checkpoint has not been evaluated for high risk or safety critical use.
+
+## License
+
+The base model `microsoft/deberta-large` is licensed under MIT.
+
+The ANLI training data is licensed under CC BY-NC 4.0.
+
+This checkpoint is released under CC BY-NC 4.0 as a conservative noncommercial choice. Users must follow the terms of the base model and the ANLI dataset.
+
+Use of this checkpoint is limited to noncommercial purposes.
+
+## Associated research
+
+This model was trained as part of the following research manuscript:
+
+**“Opening the Black Box: Localizing semantic inconsistency in NLI models with Deep k -Nearest Neighbors”**
+
+The manuscript is in preparation. It has not been submitted or published.
+
+This section will be updated when a public preprint or an accepted version becomes available.
+
+## Citation
+
+Until the paper is public, please cite this model repository:
+
+```bibtex
+@misc{mashiach2026debertaanli,
+  author = {Lidor Mashiach},
+  title = {DeBERTa Large fine tuned on ANLI},
+  year = {2026},
+  publisher = {Hugging Face},
+  url = {https://huggingface.co/Lidor-Mashiach/deberta-large-anli}
+}
+```
+
+Please also cite the DeBERTa and ANLI papers.
+
+## Contact
+
+Questions, corrections, and reproducibility reports can be posted in the Community tab of this repository.

baseline_eval.json

@@ -0,0 +1,10 @@
+{
+  "model": "DeBERTa-large",
+  "dataset": "ANLI",
+  "validation_accuracy": 0.6221875,
+  "validation_examples": 3200,
+  "validation_split": "dev",
+  "test_accuracy": 0.6178125,
+  "test_examples": 3200,
+  "test_split": "test"
+}

config.json

@@ -0,0 +1,46 @@
+{
+  "architectures": [
+    "DebertaForSequenceClassification"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "bos_token_id": null,
+  "dtype": "float32",
+  "eos_token_id": null,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 1024,
+  "id2label": {
+    "0": "entailment",
+    "1": "neutral",
+    "2": "contradiction"
+  },
+  "initializer_range": 0.02,
+  "intermediate_size": 4096,
+  "label2id": {
+    "contradiction": 2,
+    "entailment": 0,
+    "neutral": 1
+  },
+  "layer_norm_eps": 1e-07,
+  "legacy": true,
+  "max_position_embeddings": 512,
+  "max_relative_positions": -1,
+  "model_type": "deberta",
+  "num_attention_heads": 16,
+  "num_hidden_layers": 24,
+  "pad_token_id": 0,
+  "pooler_dropout": 0.0,
+  "pooler_hidden_act": "gelu",
+  "pooler_hidden_size": 1024,
+  "pos_att_type": [
+    "c2p",
+    "p2c"
+  ],
+  "position_biased_input": false,
+  "relative_attention": true,
+  "tie_word_embeddings": true,
+  "transformers_version": "5.14.1",
+  "type_vocab_size": 0,
+  "use_cache": false,
+  "vocab_size": 50265
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fbcb3fc25a88581a85e9b9491e080abc8b4bd198f32c99dde5cc3bd1f10f8dd2
+size 1624911148

model_card.json

@@ -0,0 +1,35 @@
+{
+  "model": "DeBERTa-large",
+  "dataset": "ANLI",
+  "backbone": "microsoft/deberta-large",
+  "run_seed": 1299843651,
+  "validation_accuracy": 0.6221875,
+  "training": {
+    "epochs": 3,
+    "batch_size": 16,
+    "gradient_accumulation_steps": 2,
+    "learning_rate": 1.6958369168519958e-05,
+    "weight_decay": 0.1,
+    "warmup_ratio": 0.1828387398995507,
+    "label_smoothing_factor": 0.1,
+    "adam_beta2": 0.999,
+    "adam_epsilon": 1e-08,
+    "max_grad_norm": 1.0,
+    "max_seq_len": 128,
+    "fp16": "auto",
+    "amp": "fp32"
+  },
+  "baseline_eval": {
+    "validation_accuracy": 0.6221875,
+    "test_accuracy": 0.6178125,
+    "validation_examples": 3200,
+    "test_examples": 3200,
+    "validation_split": "dev",
+    "test_split": "test"
+  },
+  "repository": "Lidor-Mashiach/deberta-large-anli",
+  "license": "cc-by-nc-4.0",
+  "base_model_license": "mit",
+  "training_data": "facebook/anli",
+  "training_data_license": "cc-by-nc-4.0"
+}

tokenizer_config.json

@@ -0,0 +1,18 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": "[CLS]",
+  "cls_token": "[CLS]",
+  "do_lower_case": false,
+  "eos_token": "[SEP]",
+  "errors": "replace",
+  "is_local": true,
+  "local_files_only": false,
+  "mask_token": "[MASK]",
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "[PAD]",
+  "sep_token": "[SEP]",
+  "tokenizer_class": "DebertaTokenizer",
+  "unk_token": "[UNK]",
+  "vocab_type": "gpt2"
+}