README.md

CodeBridge

CodeBridge is a project dedicated to optimizing large language models (LLMs) for low-resource programming languages (LRPLs) like Cangjie. It utilizes CodeBridge, a three-stage transfer learning approach, to enhance code completion accuracy by leveraging knowledge from high-resource programming languages (HRPLs) such as Java and Rust. Additionally, Retrieval-Augmented Generation (RAG) is incorporated during inference to improve performance.

Project Structure

• .env: Environment variables for API keys.
• dataset: Contains datasets for training and evaluation.
• LLaMA-Factory: Framework used for fine-tuning LLMs.
• src: Core source code directory, containing:
  • metric/: Scripts for evaluating model performance.
  • rag/: Implementations of Retrieval-Augmented Generation (RAG).
  • tree_sitter_cj/: Cangjie code parsing utilities.
  • data_cleaning.py: Scripts for preprocessing datasets.
  • inference.py: Hugging Face transformers-based inference scripts.
• inference.ipynb: Jupyter notebook for inference using vLLM.
• train.sh: Shell script for training the model.
• llm.py: Interface for interacting with the model.
• requirements.txt: Dependencies for the project.

CodeBridge: Three-Stage Fine-Tuning Process

CangjieLLM adopts CodeBridge, a novel three-stage training strategy that improves code completion for low-resource programming languages (LRPLs) through transfer learning from high-resource languages (HRPLs).

Training Strategy

1. Teaching Phase:

   • Dataset: Cangjie corpus (~8M tokens)
   • Epochs: 4
   • Learning Rate: 2e-5
   • Goal: Rapidly expose the model to Cangjie's syntax and semantics.

2. Practice Phase:

   • Dataset: Java/Rust corpus (~24M tokens)
   • Epochs: 1
   • Learning Rate: 7e-6
   • Goal: Enhance structural and semantic understanding by leveraging high-resource programming languages.

3. Correction Phase:

   • Dataset: Cangjie corpus (same as step 1)
   • Epochs: 4
   • Learning Rate: 5e-6
   • Goal: Fine-tune the model to rectify transfer-induced biases.

Dataset Preparation

• Data Sources:
  • Cangjie dataset from Huawei repositories (Cangjie-SIG, Cangjie-TPC, HW-PLLab).
  • Java/Rust dataset from StarCoder preprocessed corpus.
• Data Cleaning:
  • File filtering based on size, encoding, character composition, and comment removal.
  • Deduplication using a 90% similarity threshold.
• Data Splitting:
  • 20 projects used for evaluation (held-out test set).
  • Remaining data used for training.

Inference with RAG and Prefix Matching

For inference, Retrieval-Augmented Generation (RAG) is combined with a prefix-matching strategy to improve code completion accuracy.

Prefix-Matching Decoding Strategy

• If the input ends with a space → Extract the preceding non-space segment as prefix.
• If the input ends with a symbol → Use context-based matching to determine the appropriate completion.

This method ensures that the generated output aligns more accurately with user expectations.

Experimental Setup

• Hardware: 4x A100 GPUs (80GB)
• Batch Size: 4
• Training Time:
  • Teaching Phase: 36 hours
  • Practice Phase: 10 hours
  • Correction Phase: 36 hours

Metrics

The metric module evaluates performance using:

• Exact Match Rate (EM): Measures the percentage of perfect matches.
• Edit Similarity (ES): Computes edit distance similarity.
• Line Accuracy: Percentage of correctly generated lines within a block.

Results

1. Effectiveness of CodeBridge (RQ1)

Setting	Line-Level Exact Match Rate	Line-Level Edit Similarity	Function-Level Line Accuracy
Baseline (Untrained Model)	35.49%	0.6699	25.15%
Teaching Only	44.07%	0.7397	31.94%
No Transfer Learning (High LR)	49.44%	0.7645	30.51%
No Transfer Learning (Low LR)	46.53%	0.7568	30.70%
Transfer Learning First	47.43%	0.7563	31.91%
Full Three-Step Strategy	52.35%	0.7692	33.27%

2. Impact of Training Configurations (RQ2)

This section explores how different training settings, such as transfer data volume, final-stage learning rate, and number of epochs, influence both line-level and function-level performance.

Final Stage LR	Transfer Data Volume	Epochs	Line-Level Exact Match Rate	Line-Level Edit Similarity	Function-Level Line Accuracy
Cosine LR	-	8	45.64%	0.7432	31.78%
5e-6	1:3 (24M tokens)	4+1+4	52.35%	0.7692	33.27%
5e-6	1:1 (8M tokens)	4+1+4	40.49%	0.7165	32.00%
5e-6	1:5 (40M tokens)	4+1+4	48.32%	0.7659	32.10%
1e-5	1:3 (24M tokens)	4+1+4	47.65%	0.7660	32.70%
3e-6	1:3 (24M tokens)	4+1+4	48.77%	0.7616	32.57%
5e-6	1:3 (24M tokens)	2+1+2	48.55%	0.7497	32.16%
5e-6	1:3 (24M tokens)	3+1+3	46.98%	0.7536	32.87%

3. Generalizability of CodeBridge (RQ3)

To assess CodeBridge's generalizability, we evaluate its effectiveness across different LLM architectures and model sizes, analyzing line-level and function-level performance.

Model	Training Step	Line-Level Exact Match Rate	Line-Level Edit Similarity	Function-Level Line Accuracy
CodeLlama-13B-Instruct	Origin	32.21%	0.6695	27.24%
	Step 1	56.60%	0.8082	33.68%
	Step 2	50.56%	0.7900	32.10%
	Step 3	57.94%	0.8135	34.13%
	No Transfer	56.82%	0.8142	33.90%
Qwen2.5-14B-Instruct	Origin	29.69%	0.6342	20.60%
	Step 1	46.98%	0.7473	25.75%
	Step 2	32.89%	0.7034	21.89%
	Step 3	50.56%	0.7717	27.12%
	No Transfer	48.10%	0.7446	26.64%
DeepSeek-Coder-1.3B-Instruct	Origin	27.46%	0.5518	20.80%
	Step 1	43.53%	0.7017	26.39%
	Step 2	39.96%	0.7150	22.83%
	Step 3	44.20%	0.7179	27.28%
	No Transfer	43.97%	0.7059	26.69%
DeepSeek-Coder-6.7B-Instruct	Origin	32.37%	0.6147	23.58%
	Step 1	51.79%	0.7683	30.85%
	Step 2	39.53%	0.6546	24.60%
	Step 3	54.02%	0.7807	31.45%
	No Transfer	52.90%	0.7764	30.94%

Setup

1. Clone the repository and install dependencies:

   pip install -r requirements.txt
   cd LLaMA-Factory
   pip install -e ".[torch,metrics]"

2. Prepare the dataset and store dataset metadata in dataset_info.json.
3. Modify train.sh to specify the training configuration.
4. Run training:

   bash train.sh

5. For inference, use inference.ipynb.

Contribution

We welcome contributions! Feel free to open issues or submit pull requests.