dylanyang963/TibetanChinese-CodeMixed-OpMiner

TiChi-OpMiner: LLaMA LoRA for Tibetan–Chinese Code-Mixed Opinion Mining

TiChi-OpMiner is a LoRA-fine-tuned LLaMA model for joint sentiment and stance prediction on Tibetan–Chinese code-mixed texts.
Given a short code-mixed sentence, the model outputs:

• Sentiment: positive, neutral, or negative
• Stance: support, neutral, or oppose

The model was fine-tuned using LLaMA Factory with parameter-efficient training (LoRA).

---

1. Model Details

• Model name: your-username/tichi-opminer
• Base model: meta-llama/Llama-2-7b-chat-hf (decoder-only LLM)
• Fine-tuning method: LoRA (Low-Rank Adaptation)
• Frameworks: PyTorch, Transformers, LLaMA Factory
• Task type: joint opinion mining (sentiment + stance)
• Input languages: Tibetan (bo, Tibetan script) and Chinese (zh, simplified Chinese), with occasional English tokens / hashtags
• Output: natural language or JSON-style labels describing sentiment and stance

---

2. Intended Use

Intended use

• Research on code-mixed NLP for low-resource languages
• Experiments on Tibetan–Chinese opinion mining (social media posts, comments, short messages)
• As a starting point for further fine-tuning on related tasks (e.g., Tibetan-only or Chinese-only sentiment / stance classification)

Not intended / out of scope

• High-stakes decision making (e.g., legal, medical, financial, political decisions)
• Use on very long or domain-mismatched documents (e.g., technical reports, legal contracts)
• Any deployment scenario where incorrect predictions could cause serious harm without additional human review

---

3. Training Data

The model is trained on a 100K-instance Tibetan–Chinese code-mixed corpus:

• Each instance is a short sentence that mixes Tibetan and Chinese.
• Every sentence is annotated with:
  • Sentiment: positive, neutral, negative
  • Stance: support, neutral, oppose
• The corpus was created with a template-based generation pipeline plus manual checking to ensure:
  • Natural code-mixing patterns
  • Fluent Chinese glosses
  • Consistent joint labels
• The label space is designed so that:
  • positive ↔ support, negative ↔ oppose, neutral ↔ neutral.

If you also publish the dataset on Hugging Face, add a link here, e.g.
Dataset: your-username/tichi-opminer-dataset

---

4. Training Procedure

• Fine-tuning framework: LLaMA Factory
• Method: LoRA adapters on top of the base LLaMA model
• Objective: instruction-style text generation; the model receives a prompt containing the code-mixed text and returns both sentiment and stance.

Example (conceptual) training prompt:

You are an assistant that predicts sentiment and stance for Tibetan–Chinese code-mixed text.

Text: ང་ཚོ今天ཡིན་ནསདགའ་བསུ་བྱེད་ཡོད།
Please answer in JSON format:
{"sentiment": ..., "stance": ...}