coreason-model-foundry

Industrial Automation Engine for Training Specialized "Student Models"

The coreason-model-foundry serves as the "Refinery" in the CoReason AI ecosystem. It is an orchestrator for post-training optimization, designed to select the right mathematical strategy (DoRA, ORPO, QLoRA) for the task, prune data for maximum information density, and distribute the resulting artifacts safely.

It implements a Select-Prune-Train-Merge-Distribute Loop, utilizing unsloth for accelerated training and mergekit for model merging.

Features

• Polymorphic Training Architecture: Dynamically loads the training kernel based on the goal:
  • DoRA: Logic & Math (via UnslothSFTTrainer).
  • ORPO: Alignment & Safety (via UnslothORPOTrainer).
  • QLoRA: Memory Efficiency (via 4-bit quantization).
• Data Curator: Maximizes "Information Density" using Semantic Deduplication (SemDeDup) to remove 95%+ similar duplicates.
• Hardware Safety: "Fail Fast" mechanism prevents OOM crashes by validating VRAM requirements (e.g., enforces 24GB for full ORPO).
• The Alchemist (Merging): Integrates mergekit to combine adapters using the DARE-TIES algorithm.
• Artifact Distribution: Automatically pushes trained models to the coreason-publisher registry.
• GxP Compliance: Calculates provenance hashes (Lot Numbers) for datasets and manifests.

Installation

pip install -r requirements.txt

Note: This library relies on unsloth and torch (CUDA). Ensure these are installed in your environment suitable for your hardware.

Usage

Python API

from coreason_model_foundry import orchestrate_training

# Run the full training pipeline with a manifest file
orchestrate_training("manifest.yaml")

Example Manifest

job_id: "train-prod-2025-01-15"
base_model: "unsloth/llama-3-8b-bnb-4bit"

method_config:
  type: "orpo"
  rank: 64
  strict_hardware_check: true

dataset:
  ref: "synthesis://batch_clinical_reasoning"
  sem_dedup: true

compute:
  batch_size: 4
  grad_accum: 4