{
    "pretraining_mix": {
        "description": "Streaming blend for 1-3B salience pretraining warm-start.",
        "total_tokens_estimate": 20000000000,
        "sources": [
            {
                "name": "FineWeb-Edu",
                "dataset_id": "HuggingFaceFW/fineweb-edu",
                "config": "sample-10BT",
                "ratio": 0.6,
                "license": "CC-BY-SA-4.0 (per dataset card)",
                "streaming": true,
                "notes": "High-quality educational web subset; ~10B tokens sample."
            },
            {
                "name": "Dolma",
                "dataset_id": "allenai/dolma",
                "config": "v1_6-sample",
                "ratio": 0.3,
                "license": "ODC-BY",
                "streaming": true,
                "notes": "~10B token sample mixing web, wiki, books, code."
            },
            {
                "name": "OpenWebText",
                "dataset_id": "Skylion007/openwebtext",
                "ratio": 0.05,
                "license": "MIT",
                "streaming": true,
                "notes": "Reddit-linked open web crawl recreation."
            },
            {
                "name": "RedPajama-Data-1T-Sample",
                "dataset_id": "togethercomputer/RedPajama-Data-1T-Sample",
                "ratio": 0.05,
                "license": "Together Computer TOU",
                "streaming": true,
                "notes": "Mixed-domain slice to diversify distribution."
            }
        ]
    },
    "instruction_finetune": {
        "notes": "Select license-clean instruction corpus (e.g., OpenHermes-2.5, Open-Orca-Slim). Verify licensing before inclusion.",
        "status": "pending"
    }
}