tomyimkc/sophia-training-results

Sophia Training Results

All training history results from the sophia-agi repository.

Contents

824 result files across all Sophia four-theme training experiments:

• —Prosoche — interactive decision-making, behavioral batteries, focus experiments
• —RunPod training — LoRA training eval ladders (Qwen2.5-3B, multiple seeds/epochs)
• —Continual learning — sequential curriculum, QA judged results
• —World model — on-device, adaptive, real-corpus experiments
• —Coherence reframe — NLI gates, grounding experiments
• —Candidate integration — multi-seed A/B results
• —Format confound — constrained vs open-ended across domains
• —Scaling law — 7B → 32B → 72B results

Structure

results/
├── prosoche/           # Interactive decision-making + behavioral
├── runpod-train/       # LoRA training eval ladders
├── continual-qa/       # Continual learning
├── worldmodel-*/       # World model experiments
├── coherence-reframes/ # NLI + grounding
├── candidate-integration*/ # Multi-seed A/B
└── ...                 # 824 files total

Governance

• —canClaimAGI: false
• —candidateOnly: true
• —All results are baselines and experimental measurements — no capability claims
• —Sealed evaluation packs are barred from training

Visualization

View the interactive dashboard at tomyimkc/sophia-training-tracker.