Model save

README.md

@@ -1,87 +1,73 @@
 ---
-license: mit
 library_name: transformers
+license: apache-2.0
+base_model: distilbert-base-uncased
 tags:
-  - text-classification
-  - cryptography
-  - cryptanalysis
-  - classical-ciphers
-  - cybersecurity-education
-datasets:
-  - systemslibrarian/classical-cipher-corpus
+- generated_from_trainer
 metrics:
-  - accuracy
-  - f1
+- accuracy
+model-index:
+- name: cipher-detective-classifier
+  results: []
 ---
 
-# Cipher Detective Classifier
+<!-- This model card has been generated automatically according to the information the Trainer had access to. You
+should probably proofread and complete it, then remove this comment. -->
 
-A Transformer-based educational classifier for identifying broad classical-cipher families from ciphertext-like text.
+# cipher-detective-classifier
 
-**Part of the Cipher Detective AI project:**
-- 🕵️ Space: [systemslibrarian/cipher-detective-ai](https://huggingface.co/spaces/systemslibrarian/cipher-detective-ai)
-- 📦 Dataset: [systemslibrarian/classical-cipher-corpus](https://huggingface.co/datasets/systemslibrarian/classical-cipher-corpus)
-- 🤖 Model: [systemslibrarian/cipher-detective-classifier](https://huggingface.co/systemslibrarian/cipher-detective-classifier) _(this repo)_
+This model is a fine-tuned version of [distilbert-base-uncased](https://huggingface.co/distilbert-base-uncased) on the None dataset.
+It achieves the following results on the evaluation set:
+- Loss: 1.2801
+- Accuracy: 0.6127
+- Macro Precision: 0.6196
+- Macro Recall: 0.6392
+- Macro F1: 0.6217
 
-## Intended use
+## Model description
 
-This model supports the **Cipher Detective AI** Hugging Face Space. It is intended for:
+More information needed
 
-- cryptography education
-- classical cipher demonstrations
-- comparison against transparent heuristic baselines
-- teaching why weak ciphers leak patterns
+## Intended uses & limitations
 
-## Labels
+More information needed
 
-- `plaintext`
-- `caesar_rot`
-- `atbash`
-- `vigenere`
-- `rail_fence`
-- `columnar`
-- `affine`
-- `substitution`
+## Training and evaluation data
 
-## Not intended for
+More information needed
 
-- breaking modern encryption
-- unauthorized access
-- surveillance
-- password recovery
-- bypassing security controls
-- claims about real-world cryptanalytic capability
+## Training procedure
 
-## Training
+### Training hyperparameters
 
-```bash
-python scripts/convert_museum_corpus.py \
-  --museum /path/to/cipher-museum/public/corpus/all.jsonl \
-  --out data/cipher_examples.jsonl --split all
-python scripts/train_transformer.py \
-  --data data/cipher_examples.jsonl \
-  --model distilbert-base-uncased \
-  --out cipher_model \
-  --epochs 3 --batch-size 64
-```
+The following hyperparameters were used during training:
+- learning_rate: 2e-05
+- train_batch_size: 32
+- eval_batch_size: 32
+- seed: 42
+- gradient_accumulation_steps: 2
+- total_train_batch_size: 64
+- optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
+- lr_scheduler_type: cosine
+- lr_scheduler_warmup_steps: 0.06
+- num_epochs: 5.0
+- mixed_precision_training: Native AMP
+- label_smoothing_factor: 0.05
 
-Training data: **100,026 examples** from [systemslibrarian/classical-cipher-corpus](https://huggingface.co/datasets/systemslibrarian/classical-cipher-corpus) — sourced from the [Cipher Museum](https://ciphermuseum.com) public corpus (CC0).
+### Training results
 
-## Evaluation (3 epochs, 20% held-out test set)
+| Training Loss | Epoch | Step | Validation Loss | Accuracy | Macro Precision | Macro Recall | Macro F1 |
+|:-------------:|:-----:|:----:|:---------------:|:--------:|:---------------:|:------------:|:--------:|
+| 3.6387        | 1.0   | 912  | 1.7823          | 0.4904   | 0.5276          | 0.5280       | 0.5108   |
+| 3.0345        | 2.0   | 1824 | 1.4479          | 0.5485   | 0.5757          | 0.5806       | 0.5552   |
+| 2.7365        | 3.0   | 2736 | 1.3711          | 0.5835   | 0.6195          | 0.6139       | 0.5988   |
+| 2.5225        | 4.0   | 3648 | 1.2933          | 0.6067   | 0.6196          | 0.6332       | 0.6186   |
+| 2.6116        | 5.0   | 4560 | 1.2801          | 0.6127   | 0.6196          | 0.6392       | 0.6217   |
 
-| Metric | Value |
-|---|---|
-| Accuracy | **48.3%** |
-| Macro Precision | **56.9%** |
-| Macro Recall | **52.0%** |
-| Macro F1 | **51.8%** |
 
-81-class classification (random baseline ≈ 1.2%). Performance varies significantly by cipher family — substitution families with distinctive patterns (Caesar, Atbash, Morse) score highest; machine ciphers with similar statistical profiles (Enigma, Lorenz, SIGABA) are harder to distinguish from surface text alone.
+### Framework versions
 
-## Limitations
-
-The model learns surface statistical patterns from the ciphertext text field. It can confuse cipher families with similar statistics, especially machine ciphers, polyalphabetic variants, and short samples. It should be presented as a teaching classifier, not as a cryptanalytic authority.
-
-## Responsible-use statement
-
-Use this model to teach how classical ciphers leak patterns and why modern cryptography requires careful protocol design, key management, implementation correctness, and threat modeling.
+- Transformers 5.13.1
+- Pytorch 2.11.0+cu128
+- Datasets 4.0.0
+- Tokenizers 0.22.2