Hugging Face's logo
Akbary's picture

Akbary

akbaray
Ā·

AI & ML interests

None yet

Recent Activity

repliedto tetracta's post 17 days ago
We put Qwen2.5-3B and its Instruct fine-tune under an X-ray ā€” layer by layer, two scans running side by side on our GPU fleet. šŸ©ŗ Structural comparison (base ā†’ instruct): the change starts at layer 4 and spreads across 33 of 37 measurement points (~89% of the scanned depth) ā€” in this pair, instruction tuning is not a "last few layers" story. šŸ§  Knowledge delta, on a 20-item probe set: every fact the base model knew survived (18/20 ā†’ 18/20, zero broken), and fabrication-avoidance improved from 13/20 to 18/20. The internal true/false separation signal got slightly weaker (AUROC 0.915 ā†’ 0.878) ā€” an honest trade-off worth knowing about. Full interactive reports (no login needed): šŸ”— structure: https://www.tetracta.ai/llm_tomografi/r/a048cc12934d488aa415d0ee0c1abeca/eGWvi64CrsRQAwFirfqmVg šŸ”— knowledge: https://www.tetracta.ai/llm_tomografi/r/d0a0a6ea76f04b59b35307313b7e1cf0/WHH9qVypf4j0Fd6XPjP7Yw Model X-Ray is in free open beta ā€” 20 scans per account, any public safetensors checkpoint up to 7B, running on our own small GPU fleet: https://www.tetracta.ai/xray.html If you fine-tune or merge models: scan your checkpoint before and after, and tell us what you see ā€” we answer every message.
View all activity

Organizations

None yet