POST-TRAINING & DISTILLATION

MODULE 3: SFT DEEP DIVE · SCENE 09

Catastrophic Forgetting & Domain Adaptation

Rehearsal buffers, distillation bridges, and the retention math that keeps general capability while specializing.

FORGETTING RATEREPLAY SHAREGENERAL CAPABILITY RETAINED · 70%
retention = 1 − forgetting × (1 − replay)0%10%20%30%40%50%

Degraded: raise replay share or shorten the CPT run — the general suite will catch this in eval.

ARCHITECTURE & FORMULATIONModule 3: SFT Deep Dive

Catastrophic forgetting and the retention math

Specializing a model on new-domain data quietly erodes general capability. Continued pre-training (CPT) injects knowledge deeply but forgets fast; SFT shapes behavior gently but learns shallowly. The mitigation in both cases is rehearsal.

CPT vs SFT

CPT (raw tokens, next-token objective) moves the distribution far and risks forgetting. SFT (demonstrations) moves behavior with less collateral damage.

Replay Buffers

Mix 10–30% general-domain data into every domain-adaptation run. The model keeps rehearsing what it must not forget.

Distillation Bridges

Add a KL term pulling the fine-tuned model's outputs toward the frozen base on general prompts — a soft leash on drift.

MATHEMATICAL FORMULATION · RETENTION BOUND
retention = 1 − forgetting_rate × (1 − replay_share)

A domain run with forgetting rate 0.4 and 25% replay retains 70% of general capability; the same run without replay retains 60%.

REAL-WORLD PRODUCTION ENGINEERING
  • Medical- and legal-domain fine-tunes routinely fail public benchmarks that weren't in scope — measure retention with a frozen general eval suite.
  • A practical recipe: CPT on 1–5B domain tokens with 20% replay, then a short SFT on domain instructions with 10% general instructions.
← 08 · QLoRA Memory: 4-bit Base, Full-Pipeline Bills
10 · The Bradley-Terry Preference Model →
Agent-native web · research mode

Experimental WebMCP · evolving community-group draft, not a W3C Standard or on the Standards Track.

Native status: Initializing. Registered: none

DOM → inferred capabilities · WebMCP → declared capabilities. Tools read real CUDA/Inference Atlas content. Calls may highlight temporary UI state. No backend or remote telemetry.

Local harness verifies application logic; it does not demonstrate native discovery or a browser agent.

Input schema
{
  "type": "object",
  "properties": {
    "query": {
      "type": "string",
      "minLength": 1,
      "maxLength": 300
    },
    "limit": {
      "type": "integer",
      "minimum": 1,
      "maximum": 10
    }
  },
  "required": [
    "query"
  ],
  "additionalProperties": false
}
Local log: last 100 calls while this panel is open. Inputs may contain personal text; review before sharing.
 
[]