Self Distillation Fine Tuning SDFT: The On Policy Trick That Makes Continual Learning Finally Work

Описание: Read full article here: https://binaryverseai.com/self-distil...

Fine-tuning an LLM can feel like doing surgery with oven mitts. You ship a new skill, then discover you accidentally erased an old one. In this video, we break down Self-Distillation Fine-Tuning (SDFT), an on-policy approach that helps models keep learning without the usual catastrophic forgetting.

You’ll learn:

Why off-policy supervised fine-tuning (SFT) fails in sequential updates

How Self-Distillation uses a demo-conditioned “teacher” to correct a “student” on its own trajectories

What the results mean for continual learning, agent training, and real-world updates

When to choose weight updates vs retrieval, including LLM fine tuning vs RAG

Practical engineering details: rollouts, teacher stability, logging, and failure modes

If you’re building agents, shipping sequential model updates, or trying to add knowledge without regressions, this is the clean mental model and workflow to keep in your toolkit.

Chapters:
00:00 Intro: The On-Policy Cure
00:13 The Problem: Fine-Tuning with Oven Mitts
00:54 The Symptom: Catastrophic Forgetting
01:45 The Root Cause: Off-Policy Trajectories
03:10 The Solution: Self-Distillation Fine-Tuning (SDFT)
03:39 Methodology: Student vs. Teacher Roles
04:39 The Mechanism: Step-by-Step Correction
05:25 Analogy: The Golf Coach vs. Video
05:55 Safety Rails: Measuring Drift (Nats)
07:22 Sequential Learning: The Triple Threat Experiment
08:20 Injecting Knowledge: The 2025 Disasters Report
09:30 Comparison: SDFT vs. RAG Systems
10:35 Reasoning: Preserving the "Think" Trace
11:58 The Landscape: The Demo-Only Middle Ground
12:49 Engineering: The Three-Loop Architecture
14:02 Implementation: Teacher Stability & Logging
14:55 Philosophy: Detaching the Training Wheels
15:45 Vision: Recursive Self-Improvement
16:50 Diagnosis: When to Prescribe SDFT
17:18 Conclusion: Fix the Policy

If you found this useful, subscribe for more practical deep dives on LLM training, continual learning, and deployment tradeoffs. Drop a comment with your setup, are you doing SFT, RL, or experimenting with Self-Distillation in production?