provenance.json

{
 "version": "v0.3",
 "git_tag": "model-v0.3",
 "built_from_commit": "85956234d5e6638618b1d5c8c0907709575df3d5",
 "round": "r15-vision",
 "stage": "stage2",
 "training_commit": "aea46fc3c4bb753c13cde0a81e8261815399d9e9+dirty",
 "training_started": "2026-10-14 12:57:49",
 "stages": [
  {
   "name": "stage2",
   "out": "runs/r15-vision/stage2",
   "init": "runs/v12-ext/stage2"
  }
 ],
 "training_config": {
  "name": "r15-vision",
  "device": "mps",
  "common": {
   "log_every": 100
  },
  "stages": [
   {
    "name": "stage2",
    "kind": "train",
    "init": "runs/v12-ext/stage2",
    "data": "data/r14b/train.jsonl",
    "val": "data/r13b/val.jsonl",
    "eval": {
     "exist": "data/r13b/panel_exist.jsonl",
     "science": "data/r13b/panel_science.jsonl",
     "ext": "data/ext/dev_ext.jsonl",
     "bind_gqa": "data/r13/gqa/dev_bind_gqa.jsonl",
     "count_narrow": "data/v11/dev_count_transfer.jsonl",
     "count_wide": "data/r13b/count_wide.jsonl",
     "px_wide": "data/exp14/pixmo_count_wide.jsonl",
     "px_narrow": "data/exp14/pixmo_count_narrow.jsonl",
     "px_exist": "data/exp14/pixmo_exist.jsonl",
     "flip": "data/v6/dev_lr_pairs.jsonl",
     "vsr": "data/v5/dev_vsr.jsonl"
    },
    "params": {
     "epochs": 1,
     "lr_new": 2e-05,
     "lr_head": 1e-05,
     "lr_lora": 5e-05,
     "warmup": 300,
     "eval_every": 2000,
     "save_every": 500,
     "amp": "off",
     "seed": 1,
     "monitor_details": true,
     "vision_unfreeze_top": 4,
     "lr_vision": 5e-06,
     "start_check": "runs/exp16-r14b-s1/stage2/monitor/step0",
     "start_tolerance": 1e-05
    }
   }
  ],
  "evaluate": {
   "calibration_fit": "data/r13b/val.jsonl",
   "history": [
    "data/ext/dev_ext.jsonl",
    "data/r13/gqa/dev_bind_gqa.jsonl",
    "data/r13b/panel_exist.jsonl",
    "data/r13b/panel_science.jsonl",
    "data/r13b/count_wide.jsonl",
    "data/v11/dev_count_transfer.jsonl",
    "data/exp14/pixmo_count_wide.jsonl",
    "data/exp14/pixmo_count_narrow.jsonl",
    "data/exp14/pixmo_exist.jsonl",
    "data/v6/dev_lr_pairs.jsonl",
    "data/v5/dev_vsr.jsonl"
   ],
   "image_identity": "data/v5/image_identity.json",
   "controls": true,
   "sets": {
    "exist": {
     "path": "data/r13b/panel_exist.jsonl",
     "role": "monitoring"
    },
    "science": {
     "path": "data/r13b/panel_science.jsonl",
     "role": "monitoring"
    },
    "ext": {
     "path": "data/ext/dev_ext.jsonl",
     "role": "monitoring"
    },
    "bind_gqa": {
     "path": "data/r13/gqa/dev_bind_gqa.jsonl",
     "role": "monitoring"
    },
    "count_narrow": {
     "path": "data/v11/dev_count_transfer.jsonl",
     "role": "monitoring"
    },
    "count_wide": {
     "path": "data/r13b/count_wide.jsonl",
     "role": "monitoring"
    },
    "flip": {
     "path": "data/v6/dev_lr_pairs.jsonl",
     "role": "monitoring"
    },
    "vsr": {
     "path": "data/v5/dev_vsr.jsonl",
     "role": "monitoring"
    },
    "dev_mix": "data/v5/dev_mix.jsonl",
    "test_vsr": "data/v5/test_vsr.jsonl",
    "test_v7w": "data/v5/test_v7w.jsonl",
    "bench_pope": "data/v2/bench_pope.jsonl",
    "test_vqa_yesno": "data/v2/test_vqa_yesno.jsonl",
    "test_vqa_choice": "data/v2/test_vqa_choice.jsonl",
    "test_gqa": "data/v2/test_gqa.jsonl",
    "test_exist": "data/v2/test_exist.jsonl",
    "test_position": "data/v2/test_position.jsonl",
    "test_relation": "data/v2/test_relation.jsonl",
    "test_size": "data/v2/test_size.jsonl",
    "test_position_flip": "data/v6/test_position_flip.jsonl",
    "test_relation_flip": "data/v6/test_relation_flip.jsonl",
    "bench_lv_vqav2_yesno": "data/lv_bench/vqav2_yesno.jsonl",
    "bench_lv_aokvqa": "data/lv_bench/aokvqa.jsonl",
    "bench_lv_scienceqa": "data/lv_bench/scienceqa.jsonl",
    "bench_lv_pope": "data/lv_bench/pope.jsonl",
    "test_count_fresh": {
     "path": "data/v11/test_count_fresh.jsonl",
     "role": "heldout"
    },
    "test_bind_gqa": {
     "path": "data/r13/gqa/test_bind_gqa.jsonl",
     "role": "heldout"
    },
    "bench_countbench": {
     "path": "data/r13/bench/countbench.jsonl",
     "role": "heldout"
    },
    "bench_whatsup": {
     "path": "data/r13/bench/whatsup.jsonl",
     "role": "heldout"
    }
   },
   "commands": [
    "uv run python scripts/r13b/judge.py --rules scripts/r14b/rules.json final {eval_dir} runs/r13b-v12-base/stage2-recal || true",
    "uv run python scripts/lv_report.py {eval_dir} --prev runs/v12-ext/eval || true"
   ]
  }
 },
 "source_checkpoint_sha256": "110fea95889947c373258c88a396ad3eec574971815ee437b9c7cc3df8a5d2d5"
}