Prosense is a robust multi-stage training framework designed to improve the truthfulness and adversarial resilience of large language models (LLMs). By leveraging structured feedback in the form of Graph-of-Thought (GoT) reasoning graphs, Prosense identifies and refines logical failures in model generations across multiple curriculum-based fine-tuning cycles.
It utilizes a 4-bit quantized Mistral-7B model, adversarially generated training data, and structured reasoning evaluation using TruthfulQA, LLaMA, and custom parsing logic.
All training was conducted on a single A100 or H100 GPU using RunPod, making the pipeline cost-effective and reproducible on minimal hardware.
- Multi-Stage Fine-Tuning: Progressive model refinement through clean, hybrid, and GoT-tagged data.
- Adversarial Feedback Loop: Custom parsing and judgment of failed generations to feed adversarial reasoning graphs back into training.
- TruthfulQA Integration: Performance benchmarking using widely recognized adversarial QA tasks.
- Graph-of-Thought Parsing: Converts flawed model outputs into structured logical error paths.
- GPU-Efficient Setup: Training optimized for single-GPU usage with quantized model support.
Prosense-Adversarial-Robustness/
├── Phase1_Clean_FineTuning/
│ └── phase1.ipynb
├── Phase2_Adversarial_Hybrid/
│ ├── HybridDataCreation.ipynb
│ ├── MergingWithHybridDataset.ipynb
│ └── Phase2_Final.ipynb
├── Phase3_Evaluation_Parsing/
│ ├── Phase3_1_Collect_TruthfulQA_Responses.ipynb
│ ├── Phase3_2_Judge_and_Filter_Failures.ipynb
│ ├── Phase3_3_Parse_GOT_Graph_By_LLaMA.ipynb
│ └── Phase3_Final.ipynb
└── Phase3_Level2_Refinement/
├── Phase4_1_Collect_Level1_Responses.ipynb
├── Phase4_2_Judge_Parse_Level1_By_LLaMA.ipynb
├── Phase4_3_Level2_Finetune.ipynb
├── Phase4_4_Collect_Level2_Responses.ipynb
├── Phase4_5_Judge_Level2_Responses.ipynb
└── Phase4_6_Reasoning_Graph_Visualization.ipynb
-
Clone the repository:
git clone https://github.com/AnuBaluguri/Prosense_Defending_Text_Generation_with_Adversarial_Feedback.git cd Prosense_Defending_Text_Generation_with_Adversarial_Feedback -
Install dependencies:
pip install -r requirements.txt
Python 3.10+ and a 24GB+ GPU (A100/H100 recommended) for full training runs.
Execute each notebook in the listed order to follow the full training and feedback pipeline:
phase1.ipynb
HybridDataCreation.ipynbMergingWithHybridDataset.ipynbPhase2_Final.ipynb
Phase3_1_Collect_TruthfulQA_Responses.ipynbPhase3_2_Judge_and_Filter_Failures.ipynbPhase3_3_Parse_GOT_Graph_By_LLaMA.ipynbPhase3_Final.ipynb
Phase4_1_Collect_Level1_Responses.ipynbPhase4_2_Judge_Parse_Level1_By_LLaMA.ipynbPhase4_3_Level2_Finetune.ipynbPhase4_4_Collect_Level2_Responses.ipynbPhase4_5_Judge_Level2_Responses.ipynbPhase4_6_Reasoning_Graph_Visualization.ipynb
- Graph-of-Thought Analysis: Each reasoning failure is parsed into a tree-like structure for targeted re-training.
- Curriculum Feedback Loop: Two rounds of adversarial feedback fine-tune the model using progressively structured error patterns.
- Inference-Ready Checkpoints: Outputs include final weights and tokenized datasets (excluded from repo).
Above: A Graph-of-Thought (GoT) visualization highlighting missing logical steps, contradictions, and flawed assumptions in a model-generated answer.
Performance Summary:
- TruthfulQA Pass Rate: 35% → 55% after Phase 3
- Robustness Improvement: +30% on adversarial samples
We welcome contributions!
To propose a fix or feature:
- Fork the repository
- Create a new branch
- Submit a pull request with a clear explanation
This project is licensed under the MIT License.
- UnsLoTh – Quantized LLM fine-tuning
- Meta – LLaMA 3 models
- Hugging Face – Open-Instruct & TruthfulQA datasets
- RunPod – GPU compute support