v1.1.0

What's changed

• Manual-only RunPod deploy: Removed auto-trigger on Release; training pods launched only via workflow_dispatch with a required experiment name
• Experiment isolation: Outputs scoped to /workspace/outputs/models/<experiment> so runs don't overwrite each other
• Persistent network volume: Supports RUNPOD_VOLUME_ID for data/model persistence across pod restarts
• Training data cloning: Startup script clones elloloop/maths-questions-database via SSH deploy key; caches on volume
• Interactive pod access: SSH server + keep-alive after training for debugging and inspection
• Messages format support: Trainer handles ChatML messages format from data repo

v1.0.0

What's included

• Qwen2.5-7B LoRA fine-tuning system with GSM8K evaluation
• 3 Docker images: train, inference, eval
• Train image includes eval scripts for single-pod train+eval workflow
• Exact numeric answer matching for GSM8K scoring (local and remote)
• LoRA adapter support in BenchmarkRunner
• RunPod deployment guide with data upload instructions

v0.1.0 - Initial training system

Initial release of the Qwen2.5-7B math fine-tuning system.

What's included

• Full training pipeline with 4-phase LoRA fine-tuning
• GSM8K evaluation with error analytics
• Synthetic math data generation
• Docker image for RunPod GPU deployment
• CI/CD workflows (lint, test, release)

Docker

docker pull ghcr.io/elloloop/primary-math-finetuning:latest