Knowledge retention on Knowledge retention personalization benchmark (test)
94.3Semantic ScoreConsolidated
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| ConsolidatedLoRA adapter=Trained per conversation, Training epochs=8, Evaluation context=None2026.05 | 94.3 | 74.6 | 78.2 | 80.4 | |
| Full contextBackbone=Base model, Context=Full original conversation in system prompt, Status=ceiling2026.05 | 91.1 | 93 | 89.3 | 90.1 | |
| Compaction (C3)Summary ratio=6:1, Cycles=3, Context=Final summary as system prompt2026.05 | 62 | 36.3 | 31.5 | 36.8 | |
| No contextBackbone=Qwen2.5-7B-Instruct, Context=No system prompt and no LoRA adapter, Status=floor2026.05 | 31.8 | 9.1 | 7.7 | 11.8 |