{
  "name": "analysis/perplexity_gap/ppl_circuit_coverage/marin_32b-vs-qwen3_32b",
  "fn_name": "<function find_model_perplexity_gap at 0x7f7e9859d620>",
  "config": {
    "name": "ppl_circuit_coverage/marin_32b-vs-qwen3_32b",
    "model_a_name": "marin-community/marin-32b-base",
    "model_b_name": "Qwen/Qwen3-32B",
    "model_a_scores_path": "gs://marin-us-central1/analysis/model_perplexity_scores/ppl_circuit_coverage/marin_32b-db54dd",
    "model_b_scores_path": "gs://marin-us-central1/analysis/model_perplexity_scores/ppl_circuit_coverage/qwen3_32b-85e4f8",
    "output_path": "gs://marin-us-central1/analysis/perplexity_gap/ppl_circuit_coverage/marin_32b-vs-qwen3_32b-d0e561",
    "wandb_tags": [
      "eval=perplexity-gap",
      "dataset_bundle=ppl_circuit_coverage",
      "model_a=marin-community/marin-32b-base",
      "model_b=Qwen/Qwen3-32B"
    ],
    "retry_key": null,
    "cache_key": {
      "name": "ppl_circuit_coverage/marin_32b-vs-qwen3_32b",
      "model_a_name": "marin-community/marin-32b-base",
      "model_b_name": "Qwen/Qwen3-32B",
      "wandb_tags": [
        "eval=perplexity-gap",
        "dataset_bundle=ppl_circuit_coverage",
        "model_a=marin-community/marin-32b-base",
        "model_b=Qwen/Qwen3-32B"
      ],
      "retry_key": null
    }
  },
  "override_output_path": null,
  "version": {
    "name": "analysis/perplexity_gap/ppl_circuit_coverage/marin_32b-vs-qwen3_32b",
    "config": {
      "model_a_scores_path": "DEP[0]",
      "model_b_scores_path": "DEP[1]",
      "cache_key": {
        "name": "ppl_circuit_coverage/marin_32b-vs-qwen3_32b",
        "model_a_name": "marin-community/marin-32b-base",
        "model_b_name": "Qwen/Qwen3-32B",
        "wandb_tags": [
          "eval=perplexity-gap",
          "dataset_bundle=ppl_circuit_coverage",
          "model_a=marin-community/marin-32b-base",
          "model_b=Qwen/Qwen3-32B"
        ],
        "retry_key": null
      }
    },
    "dependencies": [
      {
        "name": "analysis/model_perplexity_scores/ppl_circuit_coverage/marin_32b",
        "config": {
          "datasets.ppl_circuit_coverage/vocab_mechanics/token_chars.input_path": "DEP[0]/vocab_mechanics_token_chars.jsonl.gz",
          "datasets.ppl_circuit_coverage/text_mechanics/character_indices.input_path": "DEP[1]/text_mechanics_character_indices.jsonl.gz",
          "datasets.ppl_circuit_coverage/text_mechanics/line_rhythm.input_path": "DEP[2]/text_mechanics_line_rhythm.jsonl.gz",
          "datasets.ppl_circuit_coverage/python_interpretation/arithmetic_repl.input_path": "DEP[3]/python_interpretation_arithmetic_repl.jsonl.gz",
          "cache_key": {
            "name": "ppl_circuit_coverage/marin_32b",
            "model": {
              "checkpoint_path": "marin-community/marin-32b-base",
              "checkpoint_is_hf": true,
              "model": null,
              "tokenizer": "marin-community/marin-tokenizer",
              "tokenizer_backend": "hf",
              "trust_remote_code": false
            },
            "datasets": {
              "ppl_circuit_coverage/vocab_mechanics/token_chars": {
                "input_path": null,
                "hf_dataset_id": null,
                "hf_dataset_name": null,
                "hf_dataset_revision": null,
                "text_key": "text",
                "input_key": "input",
                "target_key": "target",
                "split": "validation",
                "tags": [
                  "ppl_circuit_coverage",
                  "epic:5005",
                  "issue:6103",
                  "family:vocab_mechanics",
                  "task:token_chars",
                  "seed:6103",
                  "examples:tokenizer_vocab_coverage",
                  "source:generated_ppl_circuit_coverage_v1",
                  "tokenizer:marin-community/marin-tokenizer",
                  "loss:target_only",
                  "eval_only"
                ]
              },
              "ppl_circuit_coverage/text_mechanics/character_indices": {
                "input_path": null,
                "hf_dataset_id": null,
                "hf_dataset_name": null,
                "hf_dataset_revision": null,
                "text_key": "text",
                "input_key": "input",
                "target_key": "target",
                "split": "validation",
                "tags": [
                  "ppl_circuit_coverage",
                  "epic:5005",
                  "issue:6103",
                  "family:text_mechanics",
                  "task:character_indices",
                  "seed:6103",
                  "examples:1000",
                  "source:generated_ppl_circuit_coverage_v1",
                  "tokenizer:marin-community/marin-tokenizer",
                  "loss:target_only",
                  "eval_only"
                ]
              },
              "ppl_circuit_coverage/text_mechanics/line_rhythm": {
                "input_path": null,
                "hf_dataset_id": null,
                "hf_dataset_name": null,
                "hf_dataset_revision": null,
                "text_key": "text",
                "input_key": "input",
                "target_key": "target",
                "split": "validation",
                "tags": [
                  "ppl_circuit_coverage",
                  "epic:5005",
                  "issue:6103",
                  "family:text_mechanics",
                  "task:line_rhythm",
                  "seed:6103",
                  "examples:1000",
                  "source:generated_ppl_circuit_coverage_v1",
                  "tokenizer:marin-community/marin-tokenizer",
                  "loss:target_only",
                  "eval_only"
                ]
              },
              "ppl_circuit_coverage/python_interpretation/arithmetic_repl": {
                "input_path": null,
                "hf_dataset_id": null,
                "hf_dataset_name": null,
                "hf_dataset_revision": null,
                "text_key": "text",
                "input_key": "input",
                "target_key": "target",
                "split": "validation",
                "tags": [
                  "ppl_circuit_coverage",
                  "epic:5005",
                  "issue:6103",
                  "family:python_interpretation",
                  "task:arithmetic_repl",
                  "seed:6103",
                  "examples:1000",
                  "source:generated_ppl_circuit_coverage_v1",
                  "tokenizer:marin-community/marin-tokenizer",
                  "loss:target_only",
                  "eval_only"
                ]
              }
            },
            "resource_config": {
              "cpu": 32,
              "ram": "128g",
              "disk": "50g",
              "device": {
                "variant": "v5p-8",
                "kind": "tpu",
                "topology": null
              },
              "preemptible": true,
              "regions": null,
              "zone": "us-central1-a",
              "replicas": 1,
              "device_alternatives": null,
              "image": null
            },
            "per_device_batch_size": 1,
            "max_eval_length": 4096,
            "max_docs_per_dataset": null,
            "max_doc_bytes": 32768,
            "wandb_tags": [
              "eval=model-perplexity",
              "dataset_bundle=ppl_circuit_coverage",
              "model=marin-community/marin-32b-base"
            ]
          }
        },
        "dependencies": [
          {
            "name": "raw/evals/ppl_circuit_coverage",
            "config": {
              "attrs": {
                "source": "generated_ppl_circuit_coverage_v1",
                "examples_per_config": 1000,
                "seed": 6103,
                "tokenizer": "marin-community/marin-tokenizer",
                "schema": [
                  "input",
                  "target",
                  "id",
                  "subset",
                  "task",
                  "seed",
                  "metadata"
                ]
              }
            },
            "dependencies": []
          },
          {
            "name": "raw/evals/ppl_circuit_coverage",
            "config": {
              "attrs": {
                "source": "generated_ppl_circuit_coverage_v1",
                "examples_per_config": 1000,
                "seed": 6103,
                "tokenizer": "marin-community/marin-tokenizer",
                "schema": [
                  "input",
                  "target",
                  "id",
                  "subset",
                  "task",
                  "seed",
                  "metadata"
                ]
              }
            },
            "dependencies": []
          },
          {
            "name": "raw/evals/ppl_circuit_coverage",
            "config": {
              "attrs": {
                "source": "generated_ppl_circuit_coverage_v1",
                "examples_per_config": 1000,
                "seed": 6103,
                "tokenizer": "marin-community/marin-tokenizer",
                "schema": [
                  "input",
                  "target",
                  "id",
                  "subset",
                  "task",
                  "seed",
                  "metadata"
                ]
              }
            },
            "dependencies": []
          },
          {
            "name": "raw/evals/ppl_circuit_coverage",
            "config": {
              "attrs": {
                "source": "generated_ppl_circuit_coverage_v1",
                "examples_per_config": 1000,
                "seed": 6103,
                "tokenizer": "marin-community/marin-tokenizer",
                "schema": [
                  "input",
                  "target",
                  "id",
                  "subset",
                  "task",
                  "seed",
                  "metadata"
                ]
              }
            },
            "dependencies": []
          }
        ]
      },
      {
        "name": "analysis/model_perplexity_scores/ppl_circuit_coverage/qwen3_32b",
        "config": {
          "datasets.ppl_circuit_coverage/vocab_mechanics/token_chars.input_path": "DEP[0]/vocab_mechanics_token_chars.jsonl.gz",
          "datasets.ppl_circuit_coverage/text_mechanics/character_indices.input_path": "DEP[1]/text_mechanics_character_indices.jsonl.gz",
          "datasets.ppl_circuit_coverage/text_mechanics/line_rhythm.input_path": "DEP[2]/text_mechanics_line_rhythm.jsonl.gz",
          "datasets.ppl_circuit_coverage/python_interpretation/arithmetic_repl.input_path": "DEP[3]/python_interpretation_arithmetic_repl.jsonl.gz",
          "cache_key": {
            "name": "ppl_circuit_coverage/qwen3_32b",
            "model": {
              "checkpoint_path": "Qwen/Qwen3-32B",
              "checkpoint_is_hf": true,
              "model": null,
              "tokenizer": "Qwen/Qwen3-32B",
              "tokenizer_backend": "hf",
              "trust_remote_code": false
            },
            "datasets": {
              "ppl_circuit_coverage/vocab_mechanics/token_chars": {
                "input_path": null,
                "hf_dataset_id": null,
                "hf_dataset_name": null,
                "hf_dataset_revision": null,
                "text_key": "text",
                "input_key": "input",
                "target_key": "target",
                "split": "validation",
                "tags": [
                  "ppl_circuit_coverage",
                  "epic:5005",
                  "issue:6103",
                  "family:vocab_mechanics",
                  "task:token_chars",
                  "seed:6103",
                  "examples:tokenizer_vocab_coverage",
                  "source:generated_ppl_circuit_coverage_v1",
                  "tokenizer:marin-community/marin-tokenizer",
                  "loss:target_only",
                  "eval_only"
                ]
              },
              "ppl_circuit_coverage/text_mechanics/character_indices": {
                "input_path": null,
                "hf_dataset_id": null,
                "hf_dataset_name": null,
                "hf_dataset_revision": null,
                "text_key": "text",
                "input_key": "input",
                "target_key": "target",
                "split": "validation",
                "tags": [
                  "ppl_circuit_coverage",
                  "epic:5005",
                  "issue:6103",
                  "family:text_mechanics",
                  "task:character_indices",
                  "seed:6103",
                  "examples:1000",
                  "source:generated_ppl_circuit_coverage_v1",
                  "tokenizer:marin-community/marin-tokenizer",
                  "loss:target_only",
                  "eval_only"
                ]
              },
              "ppl_circuit_coverage/text_mechanics/line_rhythm": {
                "input_path": null,
                "hf_dataset_id": null,
                "hf_dataset_name": null,
                "hf_dataset_revision": null,
                "text_key": "text",
                "input_key": "input",
                "target_key": "target",
                "split": "validation",
                "tags": [
                  "ppl_circuit_coverage",
                  "epic:5005",
                  "issue:6103",
                  "family:text_mechanics",
                  "task:line_rhythm",
                  "seed:6103",
                  "examples:1000",
                  "source:generated_ppl_circuit_coverage_v1",
                  "tokenizer:marin-community/marin-tokenizer",
                  "loss:target_only",
                  "eval_only"
                ]
              },
              "ppl_circuit_coverage/python_interpretation/arithmetic_repl": {
                "input_path": null,
                "hf_dataset_id": null,
                "hf_dataset_name": null,
                "hf_dataset_revision": null,
                "text_key": "text",
                "input_key": "input",
                "target_key": "target",
                "split": "validation",
                "tags": [
                  "ppl_circuit_coverage",
                  "epic:5005",
                  "issue:6103",
                  "family:python_interpretation",
                  "task:arithmetic_repl",
                  "seed:6103",
                  "examples:1000",
                  "source:generated_ppl_circuit_coverage_v1",
                  "tokenizer:marin-community/marin-tokenizer",
                  "loss:target_only",
                  "eval_only"
                ]
              }
            },
            "resource_config": {
              "cpu": 32,
              "ram": "128g",
              "disk": "50g",
              "device": {
                "variant": "v5p-8",
                "kind": "tpu",
                "topology": null
              },
              "preemptible": true,
              "regions": null,
              "zone": "us-central1-a",
              "replicas": 1,
              "device_alternatives": null,
              "image": null
            },
            "per_device_batch_size": 1,
            "max_eval_length": 4096,
            "max_docs_per_dataset": null,
            "max_doc_bytes": 32768,
            "wandb_tags": [
              "eval=model-perplexity",
              "dataset_bundle=ppl_circuit_coverage",
              "model=Qwen/Qwen3-32B"
            ]
          }
        },
        "dependencies": [
          {
            "name": "raw/evals/ppl_circuit_coverage",
            "config": {
              "attrs": {
                "source": "generated_ppl_circuit_coverage_v1",
                "examples_per_config": 1000,
                "seed": 6103,
                "tokenizer": "marin-community/marin-tokenizer",
                "schema": [
                  "input",
                  "target",
                  "id",
                  "subset",
                  "task",
                  "seed",
                  "metadata"
                ]
              }
            },
            "dependencies": []
          },
          {
            "name": "raw/evals/ppl_circuit_coverage",
            "config": {
              "attrs": {
                "source": "generated_ppl_circuit_coverage_v1",
                "examples_per_config": 1000,
                "seed": 6103,
                "tokenizer": "marin-community/marin-tokenizer",
                "schema": [
                  "input",
                  "target",
                  "id",
                  "subset",
                  "task",
                  "seed",
                  "metadata"
                ]
              }
            },
            "dependencies": []
          },
          {
            "name": "raw/evals/ppl_circuit_coverage",
            "config": {
              "attrs": {
                "source": "generated_ppl_circuit_coverage_v1",
                "examples_per_config": 1000,
                "seed": 6103,
                "tokenizer": "marin-community/marin-tokenizer",
                "schema": [
                  "input",
                  "target",
                  "id",
                  "subset",
                  "task",
                  "seed",
                  "metadata"
                ]
              }
            },
            "dependencies": []
          },
          {
            "name": "raw/evals/ppl_circuit_coverage",
            "config": {
              "attrs": {
                "source": "generated_ppl_circuit_coverage_v1",
                "examples_per_config": 1000,
                "seed": 6103,
                "tokenizer": "marin-community/marin-tokenizer",
                "schema": [
                  "input",
                  "target",
                  "id",
                  "subset",
                  "task",
                  "seed",
                  "metadata"
                ]
              }
            },
            "dependencies": []
          }
        ]
      }
    ]
  },
  "dependencies": [
    "gs://marin-us-central1/analysis/model_perplexity_scores/ppl_circuit_coverage/marin_32b-db54dd",
    "gs://marin-us-central1/analysis/model_perplexity_scores/ppl_circuit_coverage/qwen3_32b-85e4f8"
  ],
  "output_path": "gs://marin-us-central1/analysis/perplexity_gap/ppl_circuit_coverage/marin_32b-vs-qwen3_32b-d0e561"
}
