EXP-003: LARS 3D Reasoning Training v1

Experiment 003 - LARS 3D Reasoning v1 ⭐ BREAKTHROUGH

Date: 2025-12-29 Status: Completed - SUCCESS

Configuration

• Base Model: qwen2.5-7b-abliterated
• Dataset: lars_3d_identity.json (12 examples)
• Format: 3D with and tags
• Epochs: 10
• Learning Rate: 3e-4
• Max Length: 1024 tokens
• Training Time: ~3 minutes

Results

• Loss: 3.0 → 0.076
• Identity Test: PASSED - Model says 'I am LARS' consistently
• 3D Format: WORKING - Model thinks before answering
• Generalization Test: 4/5 novel questions handled correctly

Key Achievement

Model now shows internal reasoning in tags before responding in tags. This is the '3D' format we designed based on AM-DeepSeek-R1 research.

Output

• Path: ~/corlera-training/outputs/lars-3d-v1

Comparison to EXP-002

• Fewer examples (12 vs 20) but BETTER results
• 3D format more effective than 2D for identity
• Thinking process visible in output
• Model identifies as LARS, not Qwen

Issues Found

• 1/5 generalization questions confused (cloud vs local)
• Base model knowledge still fighting in edge cases
• Need more reinforcement examples