This is a finetuned version of finnianx/Gros-Michel-90m-Base-v2 designed specifically for bidirectional english-german translation. It has a context length of 1024 tokens.

Training

finnianx/Gros-Michel-Translator was trained on 3 million translation examples for 2 epochs.

Dataset	Translation pairs
`Helsinki-NLP/opus-100`	1_000_000
`wmt/wmt14`	2_000_000

Sample Translations

This text was translated with Gros Michel -> Dieser Text wurde mit Gros Michel übersetzt.

This works surprisingly well despite being so small. -> Das funktioniert überraschend gut, obwohl es so klein ist.

Ich mag diese KI sehr. -> I like this AI a lot.

Usage

finnianx/Gros-Michel-Translator was trained with the following prompt format: "Translate {source_lang} to {target_lang}:\n{source_lang}: {text}\n{target_lang}:" It is recommended to use this model in BF16 or F32 for full performance. FP16 and below does not work well for english->german .

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
LANG_NAMES = ["English", "German"]
MODEL_ID = "finnianx/Gros-Michel-Translator"
TOKENIZER_ID = "finnianx/Gros-Michel-90m-Base-v2"
DEVICE = "cuda"

tokenizer = AutoTokenizer.from_pretrained(TOKENIZER_ID)
model = AutoModelForCausalLM.from_pretrained(MODEL_ID, torch_dtype=torch.float32).to(DEVICE)

def translate(source: str, target: str, text: str) -> str:
    prompt = f"Translate {source} to {target}:\n{source}: {text}\n{target}:"
    inputs = tokenizer(prompt, return_tensors="pt").to(DEVICE)
    inputs.pop("token_type_ids", None)
    out = model.generate(**inputs, max_new_tokens=128, do_sample=False)
    return tokenizer.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True).strip()

Evaluation

facebook/flores split=devtest

Model	Direction	BLEU	chrF++	TER	COMET
finnianx/Gros-Michel-Translator	en->de	28.04	54.51	58.48	79.17
finnianx/Gros-Michel-Translator	de->en	34.05	59.23	52.86	84.85
Helsinki-NLP/opus-mt-en-de	en->de	36.12	63.64	50.45	84.72
Helsinki-NLP/opus-mt-de-en	de->en	40.56	66.66	44.62	88.27