ilyyeees/byt5-leetspeak-decoder-v2

ByT5 Leetspeak Decoder V2

Translates leetspeak, internet slang, and gaming abbreviations back to clean English.

Built on google/byt5-base. V2 trained on real Reddit comments for improved slang handling.

Performance

Metric	V1	V2
Accuracy	71%	85%
Training Data	WikiText (synthetic)	Reddit (real)

Usage

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

model = AutoModelForSeq2SeqLM.from_pretrained("ilyyeees/byt5-leetspeak-decoder-v2")
tokenizer = AutoTokenizer.from_pretrained("ilyyeees/byt5-leetspeak-decoder-v2")

def translate(text):
    inputs = tokenizer(text, return_tensors="pt")
    outputs = model.generate(**inputs, max_length=256)
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

# Examples
print(translate("idk wh4t 2 d0 tbh"))  # I don't know what to do to be honest.
print(translate("c u l8r m8"))         # See you later mate.
print(translate("brb in 10"))          # be right back in 10
print(translate("g2g l8r m8"))         # got to go later mate
print(translate("1 h4v3 2 c4ts"))      # I have 2 cats

What It Handles

• Leetspeak: h3ll0 w0rld → hello world
• Slang: tbh, idk, rn, ngl, afk
• Gaming: gg wp, brb, g2g, 1v1
• Numbers: Preserves real numbers (2 cats stays 2 cats)
• Context: 2 late → too late vs 2 cats → 2 cats

Training

• Base: google/byt5-base (580M params)
• V1: WikiText + SAMSum + synthetic corruption
• V2: Real Reddit comments (5k) + Qwen 2.5 32B translations + continued training

Links

• GitHub
• V1 Model