Dataset	Description
techiaith/llyw-cymru-en-cy-ogl	Welsh Government / Llywodraeth Cymru parallel corpus
mgrbyte/bydtermcymru-tm-en-cy	Byd Term Cymru terminology translation memory
mgrbyte/cardiff-university-tm-en-cy	Cardiff University translation memory
mgrbyte/cwm-taf-morgannwg-university-health-board-tm-en-cy	Cwm Taf Morgannwg University Health Board translation memory
HuggingFaceFW/finetranslations (cym_Latn subset)	FineWeb-derived Welsh/English parallel sentences

All pairs were filtered to a maximum of 200 characters per side, and pairs with a length ratio greater than 3:1 were removed. Tokenised sequences longer than 128 tokens were also excluded.

Training configuration

Parameter	Value
Base model	facebook/m2m100_1.2B
Epochs	1
Batch size (per device)	32
Gradient accumulation steps	4
Learning rate	2e-5
LR scheduler	cosine
Warmup steps	6,000
Precision	fp16
Best checkpoint	step 70,000

Usage

python

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("DewiBrynJones/m2m100_1.2B-ft-cy-to-en")
model = AutoModelForSeq2SeqLM.from_pretrained("DewiBrynJones/m2m100_1.2B-ft-cy-to-en")

tokenizer.src_lang = "cy"
inputs = tokenizer("Sut mae?", return_tensors="pt")
generated = model.generate(**inputs, forced_bos_token_id=tokenizer.get_lang_id("en"))
print(tokenizer.decode(generated[0], skip_special_tokens=True))