---
library_name: pytorch
pipeline_tag: translation
tags:
- transformer
- seq2seq
- cryptanalysis
- from-scratch
- anlp
---

# C4 -- RMSNorm instead of LayerNorm

Encoder-decoder transformer written from scratch -- no `nn.Transformer`, no
pretrained weights, tokenizer fitted in-repo -- trained to decipher a binary
cipher into English plaintext. One row of a five-configuration ablation in
which exactly one architectural choice moves at a time.

| | |
| --- | --- |
| configuration | `C4` -- RMSNorm instead of LayerNorm |
| positional encoding | sinusoidal |
| attention | mha |
| normalisation | rmsnorm |
| tokenisation | unigram LM subword, fitted on the training split |
| depth | 2 encoder / 2 decoder layers |
| width | dim_model 128, 4 heads, dim_ff 512 |
| parameters | 2,761,728 |
| windowing | 1024-character source windows at stride 768 (25,654 training windows) |
| best epoch | 95 (val loss 0.0932) |

## Test metrics

| metric | test | test_full | gen_test_win | gen_test | gen_test_full |
| --- | --- | --- | --- | --- | --- |
| loss | 0.1127 | 7.0180 | -- | -- | -- |
| perplexity | 1.1193 | 1116.5630 | -- | -- | -- |
| token acc | 0.9778 | 0.2683 | 0.8427 | 0.4516 | 0.1518 |
| bit acc | -- | -- | 0.9582 | 0.8168 | 0.2638 |
| seq acc | 0.6063 | 0.0528 | 0.6996 | 0.2480 | 0.0500 |
| levenshtein | -- | -- | 0.9809 | 17.9300 | 491.9480 |
| lev / len | -- | -- | 0.0079 | 0.0263 | 0.6246 |
| BLEU | -- | -- | 0.9555 | 0.9372 | 0.1109 |
| ROUGE-1 | -- | -- | 0.9793 | 0.9688 | 0.4146 |
| ROUGE-2 | -- | -- | 0.9625 | 0.9505 | 0.3715 |
| ROUGE-L | -- | -- | 0.9793 | 0.9687 | 0.4112 |

- `test` -- teacher-forced, per window
- `test_full` -- teacher-forced, per whole line
- `gen_test_win` -- greedy, per window
- `gen_test` -- greedy, windowed and stitched
- `gen_test_full` -- greedy, whole line in one pass

## Files

- `C4_best.pt` -- weights, optimiser state, config and vocabulary sizes,
  from the best epoch.
- `config.json` -- the full ExperimentConfig this was trained under.
- `summary.json` -- the run record: the metrics above plus per-epoch history.
- `tokenizer_cipher.json`, `tokenizer_plain.json` -- the fitted unigram
  vocabularies, when the config uses them. Byte-level configurations have none
  by construction.

## Loading it

The checkpoint is a plain `torch.save` dict, not a `transformers` model. The
vocabulary sizes travel with it because they are not recoverable from the
config -- the tokenizer lands wherever pruning leaves it:

```python
import torch
from huggingface_hub import hf_hub_download

from src.config import get_config
from src.models.base import build_model

path = hf_hub_download("winterdewdev/anlp-a1-transformers-C4", "C4_best.pt")
state = torch.load(path, map_location="cpu", weights_only=False)

cfg = get_config(state["config_name"])
model = build_model(cfg, state["src_vocab_size"], state["tgt_vocab_size"])
model.load_state_dict(state["model"])
model.eval()
```

Trained with `python -m src.train --config C4`.