| """Dataset source adapters. | |
| `load_streaming_dataset` (`mindxtrain.data.curate`) dispatches to one of these | |
| based on `DataCfg.source`. Each adapter yields rows in the canonical shape | |
| expected downstream: | |
| {"messages": [{"role": "...", "content": "..."}, ...]} | |
| or a flat `{"text": "..."}` for completion-style corpora. | |
| Adapters are intentionally small and have no GPU/heavy-dep imports at module | |
| load time, so the CLI/Coach can validate configs against arbitrary sources | |
| without pulling `datasets` or `transformers`. | |
| """ | |