merging-eval 0.1.4

Model Merging Scaling Laws in Large Language Models - Empirical scaling laws for language model merging measured by cross-entropy

Merging-EVAL: Model Evaluation Framework

A comprehensive evaluation framework for language models with support for multiple datasets, GPU specification, and offline evaluation capabilities.

Features

• 🚀 Multi-dataset Support: Evaluate on code, algebra, analysis, and other domains
• 🎯 GPU Specification: Run evaluations on specific GPUs
• 📊 Offline Mode: Evaluate models without internet connection
• 🔧 Data Slicing: Evaluate on specific data subsets using indices
• 💾 Caching: Intelligent caching for faster repeated evaluations
• 📈 Comprehensive Metrics: Cross-entropy loss, token-level analysis

Repository Structure

.
├── scripts/
│   └── eval.py           # Main evaluation script
├── src/merge/
│   └── main_merging.py   # Model merging script
├── data/
│   ├── eval_partial/     # Evaluation datasets
│   │   ├── code.json     # Code generation tasks
│   │   ├── algebra.json  # Mathematical problems
│   │   └── analysis.json # Data analysis tasks
│   └── train_partial/    # Training datasets
├── test_result/          # Evaluation results and merged models
└── cache/                # Cached tokenized data

Installation and Setup

1. Environment Setup

Activate the required Python environment:

source /zju_0038/jinjia/workspace/Merging-Scaling-Law-main/merge-eval-py312/bin/activate

2. Environment Variables

Set required environment variables:

export TRANSFORMERS_NO_TORCHVISION=1
export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python

Usage

Model Merging

Merge multiple models using Task Arithmetic:

export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
python3 src/merge/main_merging.py \
  --merge_method task_arithmetic \
  --output_dir /path/to/output \
  --base_model /path/to/base/model \
  --models_to_merge "/path/to/model1,/path/to/model2,/path/to/model3" \
  --scaling_coefficient 0.2 \
  --use_gpu

GPU-Specific Merging

Run model merging on a specific GPU:

export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
export CUDA_VISIBLE_DEVICES=7
python3 src/merge/main_merging.py \
  --merge_method task_arithmetic \
  --output_dir /zju_0038/test_merge/Merging-EVAL/test_result/scheme1_16models_task_arithmetic \
  --base_model /zju_0038/wyy/mergebench/models/Llama-3.2-3B \
  --models_to_merge "/zju_0038/yifyang/scripts/models/llama-instruct-3B-v2-algebra,/zju_0038/yifyang/scripts/models/llama-instruct-3B-v2-analysis,/zju_0038/yifyang/scripts/models/llama-instruct-3B-v2-number_theory,/zju_0038/yifyang/scripts/models/llama-instruct-3B-v2-physics" \
  --scaling_coefficient 0.2 \
  --use_gpu

Basic Evaluation

Evaluate a model on a specific dataset:

python3 scripts/eval.py \
  --model /path/to/model \
  --tokenizer /path/to/tokenizer \
  --file /path/to/dataset.json \
  --output ./results \
  --batch_size 1 \
  --max_length 2048

GPU-Specific Evaluation

Run evaluation on a specific GPU:

export CUDA_VISIBLE_DEVICES=7
python3 scripts/eval.py \
  --model /path/to/model \
  --tokenizer /path/to/tokenizer \
  --file /path/to/dataset.json \
  --output ./results \
  --batch_size 1 \
  --max_length 2048 \
  --gpu_id 0 \
  --offline

Dataset-Specific Examples

Mathematical Problems (Algebra)

# 数学问题推荐 max_length = 2048
source /zju_0038/jinjia/workspace/Merging-Scaling-Law-main/merge-eval-py312/bin/activate && \
export TRANSFORMERS_NO_TORCHVISION=1 && \
export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python && \
export CUDA_VISIBLE_DEVICES=7 && \
python3 /zju_0038/test_merge/Merging-EVAL/scripts/eval.py \
  --model /zju_0038/jinjia/workspace/Merging-Scaling-Law-main/models/merged/Llama-3B-cmb/task_arithmetic_9/sc0.1_r0/6p3h \
  --tokenizer /zju_0038/jinjia/workspace/Merging-Scaling-Law-main/models/merged/Llama-3B-cmb/task_arithmetic_9/sc0.1_r0/6p3h \
  --file /zju_0038/test_merge/Merging-EVAL/data/eval_partial/algebra.json \
  --output /zju_0038/test_merge/Merging-EVAL/test_result \
  --batch_size 1 \
  --max_length 2048 \
  --gpu_id 0 \
  --offline

Code Generation Tasks

# 代码问题推荐 max_length = 4096
source /zju_0038/jinjia/workspace/Merging-Scaling-Law-main/merge-eval-py312/bin/activate && \
export TRANSFORMERS_NO_TORCHVISION=1 && \
export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python && \
export CUDA_VISIBLE_DEVICES=7 && \
python3 /zju_0038/test_merge/Merging-EVAL/scripts/eval.py \
  --model /zju_0038/jinjia/workspace/Merging-Scaling-Law-main/models/merged/Llama-3B-cmb/task_arithmetic_9/sc0.1_r0/6p3h \
  --tokenizer /zju_0038/jinjia/workspace/Merging-Scaling-Law-main/models/merged/Llama-3B-cmb/task_arithmetic_9/sc0.1_r0/6p3h \
  --file /zju_0038/test_merge/Merging-EVAL/data/eval_partial/code.json \
  --output /zju_0038/test_merge/Merging-EVAL/test_result \
  --batch_size 1 \
  --max_length 4096 \
  --gpu_id 0 \
  --offline

Parameters

Model Merging Parameters

Required Parameters

• --merge_method: Merging method (e.g., "task_arithmetic", "average_merging")
• --base_model: Path to the base model directory
• --models_to_merge: Comma-separated list of model paths to merge
• --output_dir: Output directory for merged model

Optional Parameters

• --scaling_coefficient: Scaling coefficient for merging (default: 1.0)
• --use_gpu: Use GPU for merging (default: CPU)
• --exclude_param_names_regex: Regex patterns for parameters to exclude
• --param_value_mask_rate: Parameter value mask rate (default: 0.8)
• --mask_apply_method: Method for applying masks (default: "average_merging")
• --weight_mask_rates: Comma-separated weight mask rates

Model Evaluation Parameters

Required Parameters

• --model: Path to the model directory
• --tokenizer: Path to the tokenizer directory
• --file: Path to the evaluation dataset JSON file

Optional Parameters

• --output: Output directory for results (default: ./output)
• --batch_size: Batch size for evaluation (default: 10)
• --max_length: Maximum sequence length (default: 2048)
• --gpu_id: Specific GPU ID to use (e.g., 0, 1, 2)
• --offline: Run in offline mode (no internet connection required)
• --indices: Evaluate specific data indices (e.g., "1-10,15,20-22")
• --run_name: Custom name for output folder
• --no_cache: Disable caching mechanism

Dataset-Specific Recommendations

Code Generation Tasks

• Recommended max_length: 4096-8192
• Reason: Code samples are typically longer (average: 1,899 tokens)
• Coverage: 8192 covers 98% of samples

Mathematical Problems

• Recommended max_length: 2048
• Reason: Math problems are typically shorter
• Coverage: 2048 is sufficient for most algebra problems

Output Format

Results are saved as CSV files in the output directory:

test_result/
└── model_name/
    └── all/
        └── results.csv

CSV format:

problem,CE Loss,class
algebra,0.6289,algebra
Avg.,0.6289,average
Overall,0.6289,overall

PyPI Package Distribution

This project is available as a Python package on PyPI for easy installation and distribution.

Package Information

• Package Name: merging-eval
• Version: 0.1.0
• Description: Model merging algorithms and scaling laws for large language models
• License: MIT
• Python: >=3.8

Installation

From PyPI

pip install merging-eval

With Full Dependencies

pip install "merging-eval[full]"

Development Installation

git clone https://github.com/Merging-EVAL/Merging-EVAL.git
cd Merging-EVAL
pip install -e .

Package Features

The package provides comprehensive model merging and evaluation capabilities:

import merge
from merge import MergingMethod, FlopsCounter

# Available merging methods
merging_methods = [
    "average_merging",      # Equal-weight averaging
    "task_arithmetic",      # Task vector arithmetic
    "ties_merging",         # TIES merging algorithm
    "ties_merging_dare",    # TIES with DARE variant
    "mask_merging"          # Mask-based merging
]

Publishing to PyPI

Local Development

./build_and_test.sh

Publishing

1. Setup credentials: Copy .pypirc.template to ~/.pypirc and add API tokens
2. Test PyPI: ./publish_to_pypi.sh test
3. Production PyPI: ./publish_to_pypi.sh production

Automated Publishing

• GitHub Actions automatically publishes on new releases
• Update version in pyproject.toml and create GitHub release

Troubleshooting

Model Merging Issues

1. Protobuf Version Conflicts:

   • Solution: Set export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python

2. GPU Memory Issues:

   • Use CPU mode (remove --use_gpu flag)
   • Reduce number of models to merge simultaneously
   • Use specific GPU with export CUDA_VISIBLE_DEVICES=X

3. PEFT Configuration Errors:

   • Some models may have incompatible PEFT configurations
   • Solution: Exclude problematic models or use CPU mode

4. RoPE Configuration Warnings:

   • Warning: rope_scaling configuration issues
   • Solution: These are usually non-fatal warnings

Model Evaluation Issues

1. NaN Loss Values: Usually caused by very long sequences or all-masked labels

   • Solution: Increase max_length or check data format

2. GPU Memory Issues:

   • Reduce batch_size to 1
   • Decrease max_length
   • Use specific GPU with --gpu_id

3. Network Connection Issues:

   • Use --offline flag for local model evaluation

Performance Tips

Model Merging

• Use GPU mode for faster merging when memory allows
• Start with smaller model subsets to test compatibility
• Use CPU mode for large model collections to avoid memory issues
• Set environment variables before running to avoid conflicts

Model Evaluation

• Use caching for repeated evaluations on the same dataset
• Specify GPU ID for better resource management
• Adjust max_length based on dataset characteristics
• Use offline mode when internet connection is unstable