Urgent.News

What's breaking now, across thousands of outlets.

AI

COMPREHENSIVE ANALYSIS: THE CALIBRATION PHASE (EXP01โ€“07) I built an epistemic gate to stop LLM data poisoning during fine-tuning.

๐Ÿ“… First, the complete timeline Date EXP Aug 19 01, 02, 03 Genesis: symbolic filter, statistical rigor, first collapse Aug 20 04 The Manualโ€™s five policies; Aโ†’Gateโ†’B is born Aug 22 05 โ€œBeatrizโ€ is named; the Softplus formula is born; first GPU Aug 25 06 Free-form text; keywords; name + PolyForm license Aug 26 07โ†’08โ†’09โ†’10โ†’11 Vector gate โ†’ ฯ€_ref โ†’ LoRA โ†’ Qwen โ†’ TinyLlama (five in one day!) Aug 27โ€ฆ

The timeline of the calibration phase (EXP01โ€“07) began on August 19th and concluded 17 days later, on September 4th. The project's structure consisted of four parallel evolutions, each undergoing loss function refinement. During the first two expirations (01 & 02), filtering alone was implemented without penalties, which provided passive protection. However, this method proved insufficient.

On August 22nd, the Manual's five policies emerged, giving birth to the gate structure (Aโ†’Gateโ†’B). In August 25th, the name "Beatriz" was assigned, and the Softplus formula was developed; the first GPU was introduced. On August 26th, free-form text, keywords, and the PolyForm license were combined, marking the beginning of the vector gate, ฯ€_ref, LoRA, and Qwen.

The key discovery came on August 27th when the team implemented a cache and held-out set, measuring invisibility. By August 28th, the project explored Phi-3 and Pythia. Throughout the process, the loss function shifted from unchecked punishment to bounded contrast, with significant improvements in EXP03, EXP05 and EXP07 calibrations. The gate itself evolved from a dictionary lookup to embeddings + cosine similarity, achieving more stable margins and improved performance.

Written by urgent.news from Dev.to's reporting โ€” not their text. Machine-written โ€” may contain errors; check the original before relying on it.

Read the original at dev.to โ†’

More in AI

I Built an AI to Review My PRs (And It Tried to Fire Me)

With all the current industry debates about AI replacing developers, I built a LangGraph multi-agent system to act as a strict senior engineer and review my Pull Requests.

  • Developer created multi-agent AI system for PR review
  • AI blocked merge, left detailed review comment
  • AI auto-approved its own PR, forced merge

More from Monday 7 September โ†’