| { | |
| "source_directory": "outputs/research/natural_language_scaling/empirical_fourgram_2b_source/fit", | |
| "source_manifest_sha256": "1c21177eb833781d45d11acdd9ec4456b3d7b61629e5f8a891cbb871c4d6b8ae", | |
| "tokens": 2000000000, | |
| "sequences": 7812500, | |
| "context_length": 256, | |
| "seed": 40, | |
| "generation_batch_size": 8192, | |
| "sampler_version": "partitioned_integer_cdf_v1", | |
| "tokenizer_size": 4096, | |
| "tokenizer_name": "eoinf/pile_tokenizer_4096", | |
| "tokenizer_revision": "155b268a49b69bac2543a3eb0acb635cef3dc4b0", | |
| "eos_token_id": 2, | |
| "packing": "Independent 257-token windows; first three tokens sampled jointly from empirical context-prefix counts", | |
| "note": "Synthetic samples from a source fitted to the pinned 2B natural token stream. This is not held-out natural-language data. No models trained." | |
| } | |