magic-linguistic-transfer

Plan cross-lingual adaptation of pretrained LLMs: LoRA / QLoRA / DoRA config (rank scales with typological distance, not data size), MAD-X / BAD-X adapter stacks, source-language selection via URIEL, continued-pretraining-vs-LoRA decision, catastrophic-forgetting mitigation, tool selection (Unsloth vs LLaMA-Factory vs Axolotl). Use whenever the user mentions LoRA, QLoRA, DoRA, adapter, MAD-X, BAD-X, vocab extension transfer, continued pretraining, fine-tuning a multilingual model for [language], catastrophic forgetting, Unsloth, LLaMA-Factory, Axolotl, PEFT, or asks ''how do I add [language] support to [model]''. **Use BEFORE running training jobs** — wrong adapter choice or rank wastes weeks of compute. Routed by magic-linguistic-orchestrator in the Adapt phase (now part of Acquire+Analyze).

存档时间线

版本索引永久保留；内容副本只保留最新 1 份，更早版本仅留索引与哈希（存档时间线的证据链不会因此断裂）。

下载存档内容副本