xxue752/caiti_best_model

CaiTI Best Model Bundle (Compressed Base + Task Adapters)

This repository contains the deployment-ready compressed base model and the best task-specific LoRA adapters trained for CaiTI.

Base model family

• Base architecture: meta-llama/Llama-3.2-3B-Instruct

What is included

1) Compressed base model (for adapter hot-swap deployment)

• Path: compressed_model_int4/
• Content: NF4 INT4 quantized meta-llama/Llama-3.2-3B-Instruct base model
• Intended use: low-memory deployment with task-specific adapter hot-swap

2) Task-specific adapters (recommended for max accuracy with adapter hot-swap)

• adapters/task1_response_analyzer/
  • Task: Response Analyzer (37-dimension + score classification)
• adapters/task2_general_response/
  • Task: General Response Classification (Yes/No/Maybe/Question/Stop)
• adapters/task3_rv_reasoner/
  • Task: Reflection-Validation reasoner (valid/invalid follow-up decision)
• adapters/task4_cbt_stage1/
  • Task: CBT Stage 1 (identify unhelpful thoughts)
• adapters/task4_cbt_stage2/
  • Task: CBT Stage 2 (challenge unhelpful thoughts)
• adapters/task4_cbt_stage3/
  • Task: CBT Stage 3 (reframe unhelpful thoughts)

Which adapter should be used for which task?

• Use task1_response_analyzer for dimension+score screening classification.
• Use task2_general_response for generic intent labels (Yes/No/Maybe/Question/Stop).
• Use task3_rv_reasoner for RV binary decision routing.
• Use task4_cbt_stage1, task4_cbt_stage2, and task4_cbt_stage3 for CBT multi-stage reasoning, one adapter per stage.

Deployment recommendation

• If your priority is maximum task accuracy, use task-specific adapter hot-swap.
• Recommended runtime stack: load compressed_model_int4/ once and switch adapters per task.

Notes

• These artifacts are intended for research and prototype deployment.
• Prompt design and post-processing can significantly impact output quality in generative modules.