evaluation/eight_prompt/RL/metadata.json

{
  "cell": "eval_8prompt_t1/GRPO",
  "model": "UCL-CSSB/PlasmidGPT-GRPO",
  "model_label": "GRPO",
  "model_hf_sha": "db2462aec79763c933d288165bd3c345e0f697d6",
  "timestamp_utc": "2026-05-01T23:50:12.714789Z",
  "note": "Staged from generations_sweep/temp_1.0/GRPO/outputs_vllm.csv (the 4000 seqs that produced the paper Table 1 headline of 71.6% under plasmidkit-loose QC). This copy is the input to analysis2 strict QC.",
  "source_path": "generations_sweep/temp_1.0/GRPO/outputs_vllm.csv",
  "source_metadata": {
    "model": "UCL-CSSB/PlasmidGPT-GRPO",
    "temperature": 1.0,
    "top_p": 0.9,
    "stop_token_ids": [
      2
    ],
    "samples_per_prompt": 500,
    "generator": "vllm",
    "_note": "Backfilled 2026-04-21. Original upload did not include metadata; these values are the canonical ones used for the T=1.0 GRPO evaluation (matches scripts/eval_base_sft.py sampling params and all other generations_sweep/*/GRPO/metadata_vllm.json schemas)."
  },
  "n_sequences": 4000,
  "sha256_outputs": "c68df3e4189a318abe2f61cb6925afc98201354a6de79d61a772efc30e99bbab",
  "strict_qc_analysis2": {
    "n_sequences": 4000,
    "n_passed": 2863,
    "pass_rate_pct": 71.575,
    "sha256_qc_passed": "0e999dc568ee4961603bd6b37f5cffa9659e229c8b69b9eed5db44259224bfb8",
    "sha256_qc_summary": "6aea2c1c36a341939032409c2c10f7116cf46dbfa347c70b1bacbc18a67648ca",
    "thresholds": {
      "ori_low_id": 85.0,
      "ori_low_cov": 80.0,
      "amr_low_id": 85.0,
      "amr_low_cov": 80.0,
      "ori_strict_id": 99.0,
      "ori_strict_cov": 99.0,
      "amr_strict_id": 100.0,
      "amr_strict_cov": 100.0,
      "repeat_max_len": 50
    },
    "pipeline": "analysis2: qc_oriv_arg2 + repeats2 + filter_qc_two_stage2",
    "regenerated_utc": "2026-05-01T23:54:27.703143Z"
  }
}