Qwen/Qwen2.5-0.5B-Instruct

Reproducible scope

{
  "system": {
    "id": "Qwen/Qwen2.5-0.5B-Instruct",
    "license": "apache-2.0",
    "revision": "7ae557604adf67be50417f59c2c2f167def9a775",
    "source": "https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct/tree/7ae557604adf67be50417f59c2c2f167def9a775"
  },
  "parameters": 494032768,
  "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",
    "id": "codesota-live-bpb-v1",
    "metric": "bits_per_utf8_byte",
    "runner_sha256": "db3c36e4152feb5de40ecfdb630fecc8affa4d8509612f1d4d725e16f82eeeb0",
    "selection": "20 newest cs.CL abstracts in pinned temporal manifest; domain-specific pilot, not representative general text",
    "start_token_id": 151645,
    "start_token_policy": "tokenizer BOS, else EOS; 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 0.88s · loss scoring 0.71s · peak allocated GPU memory 3.72 GiB. GPU-slot time is their sum; corpus collection and tokenization are excluded.

Document-level evidence