Neural Cipher Decryption Models (IIIT-H CSE576 ANLP Assignment 3)
Trained sequence models for character-level neural cipher decryption under long context horizons (up to 15,021 tokens), mapping an 81-character ASCII ciphertext alphabet to a 27-character plaintext alphabet (aβz and space).
Evaluated Checkpoints (Held-out Test Set, 1,318 Samples)
All models are trained with 4 layers, $d_{\text{model}}=256$, FP32:
| Model | Checkpoint File | Char Acc | Word Acc | Precision | Recall | F1 Score | Levenshtein | Params | FLOPs |
|---|---|---|---|---|---|---|---|---|---|
| DeltaNet | deltanet_best.pt |
50.40% | 45.89% | 86.43% | 73.78% | 0.7719 | 667.70 | 3,219,472 | 412,092,416 |
| Hybrid (Ours) | hybrid_best.pt |
50.03% | 45.65% | 86.21% | 73.49% | 0.7694 | 673.64 | 2,508,552 | 321,094,656 |
| Mamba | mamba_best.pt |
49.78% | 45.75% | 92.24% | 73.21% | 0.7766 | 703.08 | 1,797,632 | 230,096,896 |
| Transformer | transformer_best.pt |
44.06% | 39.29% | 86.18% | 64.07% | 0.6770 | 720.96 | 3,203,584 | 211,861,504 |
Inference Quickstart
Run evaluation or prediction directly with the repository test harness:
bash scripts/part1/test_assn.sh --model hybrid --from_hf
To test any model:
# DeltaNet
bash scripts/part1/test_assn.sh --model deltanet --from_hf
# Mamba Selective SSM
bash scripts/part1/test_assn.sh --model mamba --from_hf
# Transformer (FlashAttention SDPA)
bash scripts/part1/test_assn.sh --model transformer --from_hf
Inference Providers NEW
This model isn't deployed by any Inference Provider. π Ask for provider support