SlayerLab/Slayer149

Reproducible scope

{
  "system": {
    "adapter": "gollem",
    "id": "SlayerLab/Slayer149",
    "license": "apache-2.0",
    "revision": "7857dd28ec052cd943edcdc1f9e35f7b21dc4f3e",
    "source": "https://huggingface.co/SlayerLab/Slayer149",
    "stage": "base",
    "track": "raw-base"
  },
  "parameters": 148910738,
  "protocol": {
    "aggregation": "sum document negative log likelihood / log(2) / sum raw UTF-8 bytes",
    "batch_size": 4,
    "benchmark": "codesota-live-arxiv-2026-10-snapshot1",
    "context_tokens": 1024,
    "direction": "min",
    "dtype": "float32",
    "execution": "Native Slayer rows are unpadded and scored separately inside each batch; HF models use masked padded batches",
    "id": "codesota-live-bpb-v1",
    "loader_sources": {
      "sub1b_balanced.py": "f2949b3c2a6f3585ba9eb4b3bc7ef8306afea74492670d7c82d8c3f7e2c2a64b",
      "sub1b_gollem.py": "b9bff95fe326165b5685572727d5c50b1f86a50676932ad5b7da17018965e1d7",
      "sub1b_native.py": "ea56adc25566f7f6745ab882684dfc67e7f1cc48f29e53422fb3cab0cce769e7",
      "sub1b_slayer139.py": "f23a691bb330db4e6311b4b7e83936d5a4c584cc3f61ecf7f7025435ad87d326"
    },
    "metric": "bits_per_utf8_byte",
    "runner_sha256": "1e4ab990f08eeffdce1a6202b97b8458913510106e96fcd0b22d37cffde37065",
    "selection": "20 newest cs.CL abstracts in pinned temporal manifest; domain-specific pilot, not representative general text",
    "start_token_id": 0,
    "start_token_policy": "HF tokenizer BOS, else EOS; native config BOS or token 0; prefix not scored",
    "stride_tokens": 512,
    "task": "language-modeling",
    "tf32": false,
    "tokenization": "raw text, no chat template or added special tokens; every content token scored once"
  },
  "environment": {
    "gpu": "NVIDIA GeForce RTX 3090",
    "platform": "Linux-7.0.0-34-generic-x86_64-with-glibc2.43",
    "python": "3.11.16 (main, Sep  1 2026, 14:18:37) [Clang 22.1.3 ]",
    "torch": "2.8.0+cu128",
    "transformers": "4.56.2"
  }
}

Cost

Model loading 1.11s · loss scoring 0.33s · peak allocated GPU memory 0.84 GiB. GPU-slot time is their sum; corpus collection and tokenization are excluded.

Document-level evidence