ToneBridge-MiniCPM4.1-8B

ToneBridge-MiniCPM4.1-8B

ToneBridge-MiniCPM4.1-8B is a full merged fine-tuned model based on openbmb/MiniCPM4.1-8B.

It was created for the Hugging Face BuildSmall hackathon as the correction model powering the ToneBridge Space.

The model is designed for beginner Mandarin learners. Given a short context, target tone, and one imperfect Chinese sentence, it helps produce a more natural Chinese sentence while preserving the learner's original meaning as much as possible.

This repository contains merged model weights, not only a LoRA adapter.

What It Does

• Corrects short Chinese learner sentences.
• Fixes simple grammar, word order, word choice, and input-method mistakes.
• Adjusts tone when a sentence is inappropriate for the target context.
• Handles beginner-oriented situations such as work messages, informal team messages, friendly messages, formal WeChat messages, and informal WeChat messages.

The model should be used conservatively: if a sentence is already correct and natural enough for the selected context, the best correction may be no change.

Project Context

This model is part of ToneBridge, a Hugging Face Space built for the BuildSmall hackathon. ToneBridge focuses on helping beginner learners bridge the gap between "grammatically understandable" Chinese and Chinese that fits the intended social tone.

Training Data

The model was trained in two fine-tuning stages.

Total generated examples documented in this repository:

• 2,000 HSK 1/2 word-order examples.
• 5,000 HSK 1-3 context/tone examples.
• 7,000 total generated examples.
• 6,300 training rows used across the two stages.
• 700 validation rows used across the two stages.

The training data is synthetic and task-specific. It was generated to reduce exact duplicate source/correction pairs and to cover distinct tones and communication contexts.

The relevant JSONL training files are included in the training_data/ folder of this repository for transparency.

Training Recipe

The model was fine-tuned with a LoRA/QLoRA workflow:

1. Start from openbmb/MiniCPM4.1-8B.
2. Train a LoRA adapter on HSK 1/2 English-influenced word-order correction.
3. Continue fine-tuning that LoRA on HSK 1-3 context and tone adaptation.
4. Merge the final LoRA adapter back into the base model.
5. Upload the merged full model to this repository.

Main local artifacts:

• Stage 1 LoRA: minicpm41-hsk12-english-order-lora
• Stage 2 LoRA: minicpm41-tonebridge-v2-lora
• Merged model directory: ToneBridge-MiniCPM4.1-8B-merged

Prompt Format

The fine-tuning data used chat-style prompts. A typical training prompt looked like this:

System:
你是中文语境校对助手。只输出更合适的句子，不要解释。

User:
上下文：类别：微信非正式；场景：微信朋友聊天，轻松提醒回复；语气：短句、轻松、适合微信朋友聊天。原句的语法基本能懂，但不适合当前语境。
原句：如果你有时间，请回复我关于这个问题的消息。
任务：请根据上下文把原句改成更合适的中文。/no_think

Assistant:
有空回我一下关于这个问题的事。

Intended Use

This model is intended for:

• Mandarin learning tools.
• Beginner sentence correction.
• ToneBridge-style context-aware correction.
• Short Chinese messages where the goal is to preserve meaning while fixing grammar or tone.

It is not intended for long-form rewriting, translation, legal/medical advice, or open-ended assistant chat.

Loading Example

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Alphaplasti/ToneBridge-MiniCPM4.1-8B"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True,
    device_map="auto",
)

For memory-constrained Spaces, load the model in 4-bit quantization.

Limitations

• The training data is synthetic and narrow in scope.
• The model may over-correct sentences that are already acceptable.
• Tone adaptation can be subjective and context-dependent.
• It may not reliably handle advanced Chinese, idioms, long paragraphs, or non-beginner writing.
• For important communication, outputs should be reviewed by a human.

Base Model

This model is derived from openbmb/MiniCPM4.1-8B. Please review and respect the base model's license and usage conditions.