fix: LoRA Model Training Configuration and Data Balance

[OneZero-Y]

What type of PR is this?
Fix LoRA Model Training Configuration and Data Balance

What this PR does / why we need it:

Fixed critical configuration and data balance issues in LoRA training scripts that were causing model accuracy problems and training instability.

Key Issues Resolved

1. Enhanced Target Modules Configuration

• Improved attention learning: Added attention.self.key to BERT/RoBERTa target modules
• Better error handling: Added target modules verification and clear error messages

2. Advanced Dataset Balancing

• Intent Classification: Enhanced category balance category validation
• PII Detection: Implemented entity-type balanced sampling with negative examples
• Security Detection: Added strict 1:1 jailbreak/benign balance with distribution verification

Files Modified

• common_lora_utils.py - Core LoRA configuration and model validation
• ft_linear_lora.py - Intent classification with enhanced category balancing
• pii_bert_finetuning_lora.py - PII detection with entity-type balancing
• jailbreak_bert_finetuning_lora.py - Security detection with strict 1:1 balance

[netlify]

✅ Deploy Preview for vllm-semantic-router ready!

Name	Link
🔨 Latest commit	989b376
🔍 Latest deploy log	https://app.netlify.com/projects/vllm-semantic-router/deploys/68d7fa656fad0800083d3119
😎 Deploy Preview	https://deploy-preview-233--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/OWNER
• src/training/training_lora/classifier_model_fine_tuning_lora/ft_linear_lora.py
• src/training/training_lora/classifier_model_fine_tuning_lora/ft_linear_lora_verifier.go
• src/training/training_lora/classifier_model_fine_tuning_lora/train_cpu_optimized.sh
• src/training/training_lora/common_lora_utils.py
• src/training/training_lora/pii_model_fine_tuning_lora/pii_bert_finetuning_lora.py
• src/training/training_lora/pii_model_fine_tuning_lora/pii_bert_finetuning_lora_verifier.go
• src/training/training_lora/pii_model_fine_tuning_lora/train_cpu_optimized.sh
• src/training/training_lora/prompt_guard_fine_tuning_lora/jailbreak_bert_finetuning_lora.py
• src/training/training_lora/prompt_guard_fine_tuning_lora/jailbreak_bert_finetuning_lora_verifier.go
• src/training/training_lora/prompt_guard_fine_tuning_lora/train_cpu_optimized.sh

---

🎉 Thanks for your contributions!

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

[rootfs]

@OneZero-Y Thanks for the idea. For jailbreak 1:1 mapping, is it necessary? Would that bias towards jailbreak in evaluation?

[rootfs]

nit: can you add OWNER file in the training_lora directory and add you there?

[OneZero-Y]

Okay, I added it.

[OneZero-Y]

@rootfs
According to latest LLM Guard and Guardrails documentation :

• LLM Guard: 1:1 balanced datasets are recommended for binary security classifiers
  • Source: github.com/protectai/llm-guard/blob/main/docs/tutorials/notebooks/langchain_rag.ipynb (multi-layer detection)
  • Source: github.com/protectai/llm-guard/blob/main/docs/output_scanners/ban_topics.md (threshold tuning)
  • Source: github.com/protectai/llm-guard/blob/main/docs/output_scanners/emotion_detection.md (binary classification thresholds)
• Guardrails AI: Supports balanced training with threshold tuning for optimization
  • Source: github.com/guardrails-ai/guardrails/tree/main/docs/how_to_guides/custom_validators.md (threshold configuration)
  • Source: github.com/guardrails-ai/guardrails/tree/main/docs/examples/provenance.ipynb (validation_method="sentence")
  • Source: github.com/guardrails-ai/guardrails/tree/main/docs/examples/toxic_language.ipynb (multi-validator strategy)
• Prevent model bias toward majority class, ensure equal sensitivity to attack/benign patterns

[rootfs]

@OneZero-Y thank you for the references! They are very helpful!

[rootfs]

@OneZero-Y I am still catching up on the LoRA training. This is beyond my working experience :D

Merging the PR now