{
  "claim": "The quantization-aware training run is recorded inside the checkpoint itself. Needle 3's safetensors __metadata__ carries the full schedule \u2014 step count, seed, the bit anneal, the dataset revision and its mixture weights \u2014 so the QAT description on the page can be checked against the artifact rather than taken on trust. One thing it shows that config.json does not: 31.15% of the training mixture is speech data, while config.json lists modalities as text only.",
  "method": "Range-request the first 8 bytes for the header length, then the header itself; parse the JSON and read __metadata__. No full download needed (8,256 bytes of a 242 MB file).",
  "source": "https://huggingface.co/Cactus-Compute/needle3/blob/main/checkpoints/needle3.safetensors",
  "captured": "2026-09-18",
  "columns": [
    {
      "key": "field",
      "label": "field"
    },
    {
      "key": "value",
      "label": "value recorded in the checkpoint",
      "align": "right"
    }
  ],
  "rows": [
    {
      "field": "stage",
      "value": "qapt"
    },
    {
      "field": "training steps",
      "value": 10000
    },
    {
      "field": "seed",
      "value": "3407"
    },
    {
      "field": "bits: start \u2192 end",
      "value": "4.0 \u2192 2.0"
    },
    {
      "field": "anneal fraction",
      "value": 0.5
    },
    {
      "field": "phase B bit scheme",
      "value": "embedding=4,stack/mhc=4,default=2"
    },
    {
      "field": "knowledge-distillation weight",
      "value": 1.0
    },
    {
      "field": "base checkpoint",
      "value": "needle3_rl.pkl"
    },
    {
      "field": "tokenizer hash",
      "value": "9697fdd2fbb126dd8de173771e49d99d"
    },
    {
      "field": "dataset repo",
      "value": "Cactus-Compute/dataset"
    },
    {
      "field": "dataset revision",
      "value": "fd2ea1dac4b0\u2026"
    },
    {
      "field": "mixture \u00b7 speech-pretrain",
      "value": "10.00%"
    },
    {
      "field": "mixture \u00b7 structured-speech",
      "value": "21.15%"
    },
    {
      "field": "mixture \u00b7 structured-targeted",
      "value": "13.50%"
    },
    {
      "field": "mixture \u00b7 structured-text",
      "value": "55.35%"
    }
  ],
  "note": "Mixture weights sum to 100%. The two speech sources \u2014 speech-pretrain and structured-speech \u2014 account for 31.15% of it. config.json's extras.modalities lists [\"text\"]."
}