C5 -- token-free BLT instead of subword

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 C5 -- token-free BLT instead of subword
positional encoding sinusoidal
attention mha
normalisation layernorm
tokenisation byte-level (token-free)
depth 2 encoder / 2 decoder layers
width dim_model 128, 4 heads, dim_ff 512
parameters 2,117,376
windowing 1024-character source windows at stride 768 (25,654 training windows)
best epoch 29 (val loss 0.0131)

Test metrics

metric test test_full gen_test_win gen_test gen_test_full
loss 0.0133 5.8030 -- -- --
perplexity 1.0134 331.3012 -- -- --
token acc 0.9995 0.3228 -- -- --
bit acc -- -- 0.9997 0.9894 0.2024
seq acc 0.9575 0.0744 0.9575 0.8040 0.0680
levenshtein -- -- 0.0854 2.0480 490.0880
lev / len -- -- 0.0007 0.0037 0.6793
  • 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

  • C5_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:

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-C5", "C5_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 C5.

Downloads last month
12
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support