Grand Diomande Research · Full HTML Reader

Training Directory

``` training/ └── ircp/ ├── full_dataset/ # Full dataset training │ ├── best_model.pt # Trained model checkpoint │ ├── inferred_config.json # Model configuration │ └── [other training files] │ ├── complete_training/ # Complete training run │ ├── outputs/ # Training outputs and logs │ ├── evaluation/ # Evaluation results │ └── backups/ # Training backups └── ircp_training_backup_20250815_173556/ ```

Agents That Account for Themselves research note experiment writeup candidate score 24 .md

Full Public Reader

Training Directory

This directory contains all machine learning training artifacts, models, and evaluation results for the IRCP project.

Structure

training/
└── ircp/
    ├── full_dataset/              # Full dataset training
    │   ├── best_model.pt         # Trained model checkpoint
    │   ├── inferred_config.json   # Model configuration
    │   └── [other training files]
    │
    ├── complete_training/         # Complete training run
    │
    ├── outputs/                   # Training outputs and logs
    │
    ├── evaluation/                # Evaluation results
    │
    └── backups/                   # Training backups
        └── ircp_training_backup_20250815_173556/

Training Artifacts

### full_dataset/
Contains the production IRCP model trained on the full dataset.

Key Files:
- `best_model.pt` - PyTorch model checkpoint
- `inferred_config.json` - Model configuration
- Training hyperparameters and metrics

Used by:
- `apps/liquid-chat-backend/main.py` - Loads this model for embeddings

### complete_training/
Additional training run with complete dataset.

### outputs/
Training logs, metrics, and intermediate outputs.

### evaluation/
Evaluation results and performance metrics.

### backups/
Historical training checkpoints and backups.

Model Details

Architecture: Custom SentenceTransformer with IRCP (Inverse Ring Contextual Propagation)

Embedding Dimension: 384

Training Data: 277 conversations from Claude AI

Purpose:
- Semantic embedding generation
- Conversation similarity search
- DLM coordinate calculation

Usage

Loading the Model

python
from pathlib import Path
from ircp.models.sentence_transformer_icp import SentenceTransformerICP
import torch
import json

# Paths
PROJECT_ROOT = Path(__file__).parent.parent
model_path = PROJECT_ROOT / "training" / "ircp" / "full_dataset" / "best_model.pt"
config_path = PROJECT_ROOT / "training" / "ircp" / "full_dataset" / "inferred_config.json"

# Load configuration
with open(config_path, "r") as f:
    config = json.load(f)

# Initialize and load model
model = SentenceTransformerICP(config)
checkpoint = torch.load(model_path, map_location="cpu")
model.load_state_dict(checkpoint.get("model_state_dict", checkpoint))
model.eval()

# Generate embeddings
embedding = model.sentence_transformer.encode(["Your text here"])

Training New Models

Place new training scripts and data in this directory structure:

training/
└── ircp/
    └── your_training_run/
        ├── train.py
        ├── data/
        └── outputs/

Notes

  • Model files can be large (100MB+)
  • Ensure sufficient disk space for training
  • Backups are kept for disaster recovery
  • Evaluation results help track model performance over time

Promotion Decision

Attach run IDs, datasets, metrics, and reproduction commands.

Source Anchor

Comp-Core/backend/cc-trajectory/legacy/cc-tpo-original/cc-tpo/training/README.md

Detected Structure

Method · Evaluation · Code Anchors · Architecture