WiNGPT-Babel-2.2

WiNGPT-Babel-2: A Multilingual Translation Language Model

Experimental updates:

1. Feels ok in use, spBLEU looks low: That’s normal — the model isn’t necessarily worse. spBLEU likes word-for-word matches; good rewrites score poorly. The 22.70 xx→zh score mixes many languages; for en↔zh, look at en→zh (36.97) instead. FLORES test text also differs from real app usage. More real-world testing is still welcome.

2. FLORES (spBLEU):

   Direction	Score
   zh → en	32.29
   xx → en	31.61
   en → zh	36.97
   xx → zh	22.70

3. Base: tencent/Hy-MT2-1.8B (~1.8B) — faster and lighter than the ~2B WiNGPT-Babel-2 backbone.

---

WiNGPT-Babel-2 is a language model optimized for multilingual translation tasks. As an iteration of WiNGPT-Babel, it features significant improvements in language coverage, data format handling, and translation accuracy for complex content.

The model continues the "Human-in-the-loop" training strategy, iteratively optimizing through the analysis of log data from real-world application scenarios to ensure its effectiveness and reliability in practical use.

Core Improvements in Version 2.0

WiNGPT-Babel-2 introduces the following key technical upgrades over its predecessor:

1. Expanded Language Support: Through training with the wmt24pp dataset, language support has been extended to 55 languages, primarily enhancing translation capabilities from English (en) to other target languages (xx).

2. Enhanced Chinese Translation: The translation pipeline from other source languages to Chinese (xx → zh) has been specifically optimized, improving the accuracy and fluency of the results.

3. Structured Data Translation: The model can now identify and translate text fields embedded within structured data (e.g., JSON) while preserving the original data structure. This feature is suitable for scenarios such as API internationalization and multilingual dataset preprocessing.

4. Mixed-Content Handling: Its ability to handle mixed-content text has been improved, enabling more accurate translation of paragraphs containing mathematical expressions (LaTeX), code snippets, and web markup (HTML/Markdown), while preserving the format and integrity of these non-translatable elements.

Training Methodology

The performance improvements in WiNGPT-Babel-2 are attributed to a continuous, data-driven, iterative training process:

1. Data Collection: Collecting anonymous, real-world translation task logs from integrated applications (e.g., Immersive Translate, Videolingo).
2. Data Refinement: Using a reward model for rejection sampling on the collected data, supplemented by manual review, to filter high-quality, high-value samples for constructing new training datasets.
3. Iterative Retraining: Using the refined data for the model's incremental training, continuously improving its performance in specific domains and scenarios through a cyclical iterative process.

Technical Specifications

• Base Model: GemmaX2-28-2B-Pretrain
• Primary Training Data: "Human-in-the-loop" in-house dataset, WMT24++ dataset
• Maximum Context Length: 4096 tokens
• Chat Capability: Supports multi-turn dialogue, allowing for contextual follow-up and translation refinement.

Language Support

Performance

<table> <thead> <tr> <th rowspan="2" align="center">Model</th> <th colspan="2" align="center">FLORES-200</th> </tr> <tr> <th align="center">xx → en</th> <th align="center">xx → zh</th> </tr> </thead> <tbody> <tr> <td align="center">WiNGPT-Babel-AWQ</td> <td align="center">33.91</td> <td align="center">17.29</td> </tr> <tr> <td align="center">WiNGPT-Babel-2-AWQ</td> <td align="center">36.43</td> <td align="center">24.45</td> </tr> </tbody> </table>

Note:

1. The evaluation metric is spBLEU, using the FLORES-200 tokenizer.

2. 'xx' represents the 52 source languages from the wmt24pp dataset.

Usage Guide

For optimal inference performance, it is recommended to use frameworks such as vllm. The following provides a basic usage example using the Hugging Face transformers library.

System Prompt: For optimal automatic language inference, it is recommended to use the unified system prompt: Translate this to {{to}} Language. Replace {{to}} with the name of the target language. For instance, use Translate this to Simplified Chinese Language to translate into Chinese, or Translate this to English Language to translate into English. This method provides precise control over the translation direction and yields the most reliable results.

Example

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "winninghealth/WiNGPT-Babel-2"

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

# Example: Translation of text within a JSON object to Chinese
prompt_json = """{
  "product_name": "High-Performance Laptop",
  "features": ["Fast Processor", "Long Battery Life", "Lightweight Design"]
}"""

messages = [
   {"role": "system", "content": "Translate this to Simplified Chinese Language"}, 
   {"role": "user", "content": prompt_json} # Replace with the desired prompt
]

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=4096,
   temperature=0
)

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]

For additional usage demos, you can refer to the original WiNGPT-Babel.

LICENSE

1. This project's license agreement is the Apache License 2.0

2. Please cite this project when using its model weights: https://huggingface.co/winninghealth/WiNGPT-Babel-2

3. Comply with gemma-2-2b, GemmaX2-28-2B-v0.1, immersive-translate, VideoLingo protocols and licenses, details on their website.

Contact Us

• Apply for a token through the WiNGPT platform https://wingpt.winning.com.cn/platform
• Contact us at [email protected]