feat: add scripts to fine tune qwen3 for knowledge specialization

[rootfs]

What type of PR is this?

Fine tune qwen3 0.6B for specialization. There are two ways: with MMLU as training dataset (i.e. leakage) for demo purpose, and use other dataset for training (no leakage).

What this PR does / why we need it:

Which issue(s) this PR fixes:

Fixes #239

Release Notes: Yes/No

[netlify]

✅ Deploy Preview for vllm-semantic-router ready!

Name	Link
🔨 Latest commit	01e9b4b
🔍 Latest deploy log	https://app.netlify.com/projects/vllm-semantic-router/deploys/68f01e7b825c8a0008c5ad5e
😎 Deploy Preview	https://deploy-preview-447--vllm-semantic-router.netlify.app
📱 Preview on mobile	Toggle QR Code... Use your smartphone camera to open QR code link.

---

To edit notification comments on pull requests, go to your Netlify project configuration.

[github-actions]

👥 vLLM Semantic Team Notification

The following members have been identified for the changed files in this PR and have been automatically assigned:

📁 src

Owners: @rootfs, @Xunzhuo, @wangchen615
Files changed:

• src/training/training_lora/mmlu_pro_solver_lora/ft_qwen3_mmlu_solver_lora.py
• src/training/training_lora/mmlu_pro_solver_lora/ft_qwen3_mmlu_solver_lora_no_leakage.py
• src/training/training_lora/mmlu_pro_solver_lora/train_all_specialists.sh
• src/training/training_lora/mmlu_pro_solver_lora/train_all_specialists_no_leakage.sh
• src/training/training_lora/README.md
• src/training/training_lora/classifier_model_fine_tuning_lora/ft_qwen3_generative_lora.py

📁 Root Directory

Owners: @rootfs, @Xunzhuo
Files changed:

• examples/mcp-classifier-server/server_generative.py

---

🎉 Thanks for your contributions!

This comment was automatically generated based on the OWNER files in the repository.

[Copilot]

Pull Request Overview

This PR adds scripts and implementations to fine-tune Qwen3 0.6B models for specialized knowledge domains using MMLU-Pro benchmarks. The key focus is on training domain-specific expert models (math, science, law, humanities, etc.) with and without data leakage scenarios to demonstrate proper evaluation methodology.

Key changes:

• Added complete training pipeline for 6 specialized Qwen3 models
• Implemented both "leakage" (demo) and "no-leakage" (proper) training approaches
• Included batch training scripts with comprehensive logging and result tracking

Reviewed Changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
train_all_specialists_no_leakage.sh	Bash script for batch training 6 specialists using external datasets (no MMLU-Pro leakage)
train_all_specialists.sh	Bash script for batch training 6 specialists using MMLU-Pro data (with leakage for demo)
ft_qwen3_mmlu_solver_lora_no_leakage.py	Python training script using external datasets (GSM8K, MATH, ARC, etc.) for proper evaluation
ft_qwen3_mmlu_solver_lora.py	Python training script using MMLU-Pro data directly (demo/leakage version)
README.md	Updated documentation to include the new MMLU-Pro solver task

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

[Copilot]

FAILED_MODELS is initialized as a scalar (=0) but later used as both an array (FAILED_MODELS+=(...)) and incremented as a counter. This will cause incorrect behavior. Use separate variables for the counter and the array.

Suggested change

	FAILED_MODELS=0
	FAILED_MODELS_COUNT=0

[rootfs]

With the no_leakage training, the math accuracy is improved after fine tuning:

train_all_specialists_no_leakage.sh 2 100 5

2025-10-15 20:10:22 - common_lora_utils - INFO - IMPROVEMENT ANALYSIS (No Data Leakage)
2025-10-15 20:10:22 - common_lora_utils - INFO - 📊📊📊📊📊📊📊📊📊📊📊📊📊📊📊📊📊📊📊📊📊📊📊📊📊📊📊📊📊📊📊📊📊📊📊📊📊📊📊📊
2025-10-15 20:10:22 - common_lora_utils - INFO - 
================================================================================
2025-10-15 20:10:22 - common_lora_utils - INFO - OVERALL RESULTS:
2025-10-15 20:10:22 - common_lora_utils - INFO - ================================================================================
2025-10-15 20:10:22 - common_lora_utils - INFO -   Baseline (Untrained):     10.00%
2025-10-15 20:10:22 - common_lora_utils - INFO -   Post-training:            24.00%
2025-10-15 20:10:22 - common_lora_utils - INFO -   Absolute Improvement:     +14.00%
2025-10-15 20:10:22 - common_lora_utils - INFO -   Relative Improvement:     +140.0%