Archive verified research artifact: effective_config.json

artifacts/experiments/runs/gate-search-mira-mmqa_ti-transfer_p2_r1-s17/effective_config.json

@@ -0,0 +1,121 @@
+{
+  "adaptations": [
+    "Qwen3-8B frozen decoder; SigLIP2 vision; BGE text embeddings",
+    "Shared Qwen3-VL source extraction plus Qwen3-8B confidence sidecar",
+    "Dataset candidate pools; no external prior KG",
+    "Canonical normalized string map; no unverified coreference resolution",
+    "ScienceQA protected question image for SCOPE source selection",
+    "FP32 evaluation, TF32 disabled; custom MMQA resplit",
+    "Shared question-specific candidate graph filtered by the MIRA confidence policy before textual-graph baselines; not the original corpus-wide benchmark graph",
+    "Separate QA training per dataset; shared visual foundation checkpoint is declared external/common training for this comparison",
+    "MIRA Eq.25 text-prefix K/V are independent linear adapters, distinct from HieraSlot Eq.12/13 shared slot adapters. Dimensions map frozen 768-d embeddings to Qwen hidden size then native GQA.",
+    "Gold evidence with no retained matching slot contributes no contrastive positive and beta=0; no replacement by a pseudo-positive."
+  ],
+  "aliases": {},
+  "blockers": [],
+  "checkpoint_selection": "Separate training per dataset; ScienceQA accuracy; MMQA list F1, tie list EM; tie earlier epoch; val_selection only",
+  "compiler_identity": "source-closed-confidence-siglip-v1",
+  "confidence": {
+    "alpha": 0.5,
+    "conflict_weight": 0.2,
+    "threshold": 0.5
+  },
+  "confidence_coefficient_origin": "Declared reconstruction choices, not published numerical defaults; tune only on development before freeze.",
+  "conflict_rules": {
+    "exclusive_objects": {},
+    "relation_negations": [
+      [
+        "is",
+        "is not"
+      ],
+      [
+        "has",
+        "does not have"
+      ],
+      [
+        "can",
+        "cannot"
+      ]
+    ]
+  },
+  "counts": {
+    "mmqa_ti": {
+      "train": 10645,
+      "val_selection": 1065
+    },
+    "scienceqa": {
+      "train": 12726,
+      "val_selection": 3393
+    }
+  },
+  "datasets": [
+    "mmqa_ti"
+  ],
+  "decoder": {
+    "do_sample": false,
+    "dtype": "bfloat16",
+    "enable_thinking": false,
+    "max_answer_tokens": 256,
+    "max_new_tokens": 256,
+    "max_prompt_tokens": 512,
+    "prefix_native_input_rmsnorm": true,
+    "reserved_prefix_positions": 256
+  },
+  "exact_original_replication": false,
+  "external_knowledge": "none-v1 (controlled adaptation)",
+  "fidelity_label": "controlled adaptation; not native exact reproduction",
+  "gate": {
+    "equation": "paper Eq.21 additive, no interaction substitution",
+    "threshold": 0.6,
+    "top_k": 16
+  },
+  "implementation_status": "implemented",
+  "method": "mira",
+  "relation_aliases": {},
+  "retrieval": {
+    "dense_top_k": 4,
+    "edge_base_cost": 0.15,
+    "edge_degree_cost": 0.02,
+    "edge_length_cost": 0.001,
+    "edge_offtopic_cost": 0.5,
+    "max_hops": 2,
+    "max_triples_per_source": 4,
+    "node_relation_weight": 0.5,
+    "node_text_weight": 1.0,
+    "offtopic_threshold": 0.5,
+    "pcst_max_edges": 24,
+    "reranker_top_k": 4
+  },
+  "run_scope": "development only; full GPU integration and benchmark quality pending",
+  "seeds": [
+    17,
+    42,
+    2026
+  ],
+  "selector": {
+    "budget": 0.75,
+    "conflict_penalty": 0.1,
+    "cost_penalty": 0.05
+  },
+  "training": {
+    "beta_max": 0.2,
+    "epochs": 3,
+    "gate_learning_rate": 1e-06,
+    "grad_clip": 1.0,
+    "gradient_accumulation": 4,
+    "lambda_gate": 0.0,
+    "lambda_gen": 0.6,
+    "lambda_retr": 0.4,
+    "learning_rate": 3e-05,
+    "temperature_nce": 0.07,
+    "warmup_epochs": 0,
+    "weight_decay": 0.01
+  },
+  "version": "methods-v1",
+  "vision": {
+    "all_candidate_images": true,
+    "max_image_tokens": 64,
+    "pooled_grid": 2
+  },
+  "visual_adapter": "experiments/runs/foundation/adapter.pt"
+}