Knowledge note
Fine-Tuning and Adaptation
Fine-tuning, transfer learning, and self-evolution training methods for language models and coding agents.
Fine-Tuning and Adaptation
- Intrinsic Dimensionality and Language Model Fine-Tuning — Intrinsic-dimension view of pretrained language model fine-tuning: DID/SAID subspace training, low
d90task dimensions, pretraining as downstream task compression, model-size trends, generalization correlations, and an editable Draw.io explainer. - Socratic-SWE: Self-Evolving Coding Agents via Trace-Derived Skills — Closed-loop self-evolution framework: trace-derived Agent Skill Registry, skill-guided Generator with four-stage Verifier Gate, gradient-aligned Generator reward, GDPO-normalized Solver reward, and 50.40% on SWE-bench Verified across three iterations.