Training/Training_Documentation.txt

ClinicalThought-AI-8B Training Documentation
===============================================

Model Training Details
---------------------

Base Model:               Granite 3.3 8B Instruct
Fine-tuning Method:       LoRA (Low-Rank Adaptation)
Training Infrastructure:  Single NVIDIA RTX 6000 Ada Generation GPU
Training Duration:        Approximately 75.8 hours
Training Dataset:         Custom curated dataset for medical reasoning

Dataset Specifications
---------------------

Total Token Count:        38,514,400
Total Sample Count:       29,500
Average Tokens/Sample:    1305.57
Dataset Creation:         Created from a combination of public medical reasoning datasets from OpenAI o1 and DeepSeek-R1, along with additional reasoning chains created using Claude Sonnet 4 extended thinking

Training Configuration
---------------------

LoRA Parameters:
- Rank:                   32
- Alpha:                  64
- Dropout:                0.1
- Target Modules:         q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj, lm_head

Training Hyperparameters:
- Learning Rate:          2e-5
- Batch Size:             1
- Gradient Accumulation:  8
- Effective Batch Size:   8
- Max Sequence Length:    12,000
- Epochs:                 8
- Warmup Ratio:           0.05
- Weight Decay:           0.005
- Max Grad Norm:          1.0
- LR Scheduler:           Cosine with Restarts

Hardware & Environment
---------------------

GPU:                      NVIDIA RTX 6000 Ada Generation (48GB)
Operating System:         Ubuntu
CUDA Version:             11.8
PyTorch Version:          2.7.0
Compute Capability:       8.9
Optimization:             FP16, Gradient Checkpointing

Training Performance
---------------------

Training Runtime:         75.8 hours (272,919 seconds)
Train Samples/Second:     0.865
Train Steps/Second:       0.108
Training Loss (Final):    0.738
Total Training Steps:     29,504