README.md

metadata

datasets:
  - lmg-anon/VNTL-v3.1-1k
language:
  - en
  - ja
base_model:
  - Qwen/Qwen2.5-7B
pipeline_tag: translation
tags:
  - translation
  - vntl
  - qwen
license: apache-2.0
library_name: transformers

Qwen2.5-7B-VNTL-JP-EN

Qwen2.5-7B finetuned for Japanese to English translation.

Trained on ~150k sentences from VNTL-v3.1-1k.

The model was trained on just the sentences in random order to make it more flexible and useful outside of just VN translation.

Usage

Ollama

1. ollama run technobyte/Qwen2.5-7B-VNTL-JP-EN:q4_k_m
2. Input just the Japanese sentence.

Llama.cpp

1. Download the GGUF.
2. llama-cli -m Qwen2.5-7B-VNTL-JP-EN-Q4_K_M.gguf -no-cnv -p "A Japanese sentence along with a proper English equivalent.\nJapanese: 放課後はマンガ喫茶でまったり〜♡ おすすめのマンガ教えて！ \nEnglish: "

Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "TechnoByte/Qwen2.5-7B-VNTL-JP-EN"

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

messages = [
    {"role": "user", "content": "放課後はマンガ喫茶でまったり〜♡ おすすめのマンガ教えて！"}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=512
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]

Prompt template

Plaintext

A Japanese sentence along with a proper English equivalent.
Japanese: JAPANESE SENTENCE HERE
English: 

Jinja (HF Transformers)

{% for i in range(0, messages|length, 2) %}A Japanese sentence along with a proper English equivalent.
Japanese: {{ messages[i].content }}
English:{% if i+1 < messages|length %} {{ messages[i+1].content }}<|endoftext|>{{ "
" }}{% else %}{% endif %}{% endfor %}

Go (Ollama)

A Japanese sentence along with a proper English equivalent.
Japanese: {{ .Prompt }}
English: {{ .Response }}<|endoftext|>

Limitations

• Can only translate one sentence per turn.
• Can use incorrect pronouns due to lack of context.