jumafernandez/trace-dialog-chain-checkpoints

trace-dialog continued-pretraining chain

Checkpoints of a dialogue-level encoder trained over frozen dialog2flow-joint-bert-base turn embeddings, continued-pretrained sequentially over large conversational corpora (see the d2f-turn-embeddings-* datasets on this profile). One subfolder per stage; each contains the best checkpoint, its training log and config. Research artifact; documentation forthcoming.

Which checkpoint should I use?

xlc-s2 — the final model of the chain. It has seen the full curriculum (~26M turns) and is the recommended checkpoint. The intermediate stages are released for research on continued-pretraining dynamics:

Stage	Trained on (cumulative)
xlc-s1	+ SODA	intermediate
xlc-s1u	+ UltraChat	intermediate
xlc-s1w	+ WildChat	intermediate
xlc-s1p	+ PersonaChat	intermediate
xlc-s2	+ Taskmaster (full curriculum)	final — use this one