candle-mi 0.1.13

Mechanistic interpretability for language models in Rust, built on candle
Documentation
{
  "model_repo": "medmekk/Llama-3.2-1B-Instruct-bnb-nf4",
  "tokenizer_repo": "meta-llama/Llama-3.2-1B",
  "methodology": "bitsandbytes-NF4 forward-pass oracle (transformers + bitsandbytes, F32 compute on CUDA); validates candle-mi's anamnesis NF4 dequant load path",
  "torch_version": "2.10.0+cu130",
  "transformers_version": "5.1.0",
  "bitsandbytes_version": "0.49.1",
  "platform": "Windows-11-10.0.26200-SP0",
  "model_type": "llama",
  "quant_method": "QuantizationMethod.BITS_AND_BYTES",
  "hidden_size": 2048,
  "num_layers": 16,
  "vocab_size": 128256,
  "head_dim": 64,
  "test_cases": [
    {
      "prompt": "The capital of France is",
      "tokens": [
        128000,
        791,
        6864,
        315,
        9822,
        374
      ],
      "top_10": [
        {
          "index": 12366,
          "logit": 19.231300354003906
        },
        {
          "index": 539,
          "logit": 16.053695678710938
        },
        {
          "index": 264,
          "logit": 15.744446754455566
        },
        {
          "index": 1131,
          "logit": 15.438555717468262
        },
        {
          "index": 279,
          "logit": 14.940621376037598
        },
        {
          "index": 20437,
          "logit": 14.774496078491211
        },
        {
          "index": 25,
          "logit": 14.530152320861816
        },
        {
          "index": 311,
          "logit": 14.517041206359863
        },
        {
          "index": 7559,
          "logit": 14.354340553283691
        },
        {
          "index": 3604,
          "logit": 14.301971435546875
        }
      ]
    },
    {
      "prompt": "Two plus two equals",
      "tokens": [
        128000,
        11874,
        5636,
        1403,
        17239
      ],
      "top_10": [
        {
          "index": 3116,
          "logit": 20.51384925842285
        },
        {
          "index": 4330,
          "logit": 20.301528930664062
        },
        {
          "index": 220,
          "logit": 18.263551712036133
        },
        {
          "index": 4400,
          "logit": 18.191823959350586
        },
        {
          "index": 8223,
          "logit": 17.986309051513672
        },
        {
          "index": 1131,
          "logit": 17.86895179748535
        },
        {
          "index": 8254,
          "logit": 17.833688735961914
        },
        {
          "index": 4848,
          "logit": 17.661401748657227
        },
        {
          "index": 198,
          "logit": 17.351215362548828
        },
        {
          "index": 482,
          "logit": 17.265073776245117
        }
      ]
    },
    {
      "prompt": "Once upon a time, there was a",
      "tokens": [
        128000,
        12805,
        5304,
        264,
        892,
        11,
        1070,
        574,
        264
      ],
      "top_10": [
        {
          "index": 3995,
          "logit": 19.341646194458008
        },
        {
          "index": 2678,
          "logit": 17.349872589111328
        },
        {
          "index": 893,
          "logit": 17.160659790039062
        },
        {
          "index": 6366,
          "logit": 17.104965209960938
        },
        {
          "index": 2697,
          "logit": 16.60169219970703
        },
        {
          "index": 8334,
          "logit": 16.12689208984375
        },
        {
          "index": 3828,
          "logit": 16.021955490112305
        },
        {
          "index": 1912,
          "logit": 15.974161148071289
        },
        {
          "index": 24632,
          "logit": 15.948994636535645
        },
        {
          "index": 5333,
          "logit": 15.813974380493164
        }
      ]
    }
  ]
}