MACE-Bench / code /experiments /retriever_pilot /plan /out /generated_weights.json
tirtho149's picture
Add retriever study (code, predictions, one-pager), refresh response cache
874eec2 verified
Raw History Blame Contribute Delete
10.5 kB
{
"generator": "sonnet (claude-sonnet-5-5), temperature default, one call per condition and pack",
"scale": {
"mu": [
0.12935667071142096,
0.2550301085011053,
0.3469950335581999,
0.8692705717262315,
-0.008899994300391548
],
"sd": [
0.07783956971687504,
0.07069942425154609,
0.05872549652863431,
0.04915714657502629,
0.27735868371550865
]
},
"runs": {
"G0_p0": {
"ok": true,
"note": "Rewards IDF coverage most since it captures discriminative terms, plus Jaccard, trigram and embedding similarity. A fourth hidden unit penalizes generic similarity and length mismatch, giving a mild penalty to distractors sharing only crop vocabulary.",
"cost_usd": 0.005186,
"cache_hit": true,
"params": {
"W1": [
[
1.5,
0.8,
0.5,
0.6,
0.0
],
[
0.9,
1.2,
0.3,
0.2,
-0.2
],
[
1.2,
-0.3,
0.8,
1.0,
0.0
],
[
-0.8,
-0.4,
-0.3,
-0.5,
0.3
]
],
"b1": [
0.0,
0.0,
0.1,
0.2
],
"w2": [
1.0,
0.6,
0.7,
-0.4
],
"b2": [
0.0
]
}
},
"G0_p1": {
"ok": true,
"note": "Scores rise with IDF coverage, which favors discriminative disease terms over shared crop vocabulary, plus Jaccard overlap, trigram and embedding similarity. A small penalty on length mismatch. Four ReLU units give a monotone, mostly additive ranking with some nonlinearity.",
"cost_usd": 0.005126,
"cache_hit": true,
"params": {
"W1": [
[
1.5,
0.5,
0.3,
0.5,
0.0
],
[
0.8,
1.2,
0.5,
0.3,
-0.3
],
[
1.2,
-0.3,
0.2,
0.8,
0.0
],
[
0.3,
0.3,
1.0,
0.2,
-0.2
]
],
"b1": [
0.0,
0.0,
0.0,
0.0
],
"w2": [
0.9,
0.5,
0.6,
0.3
],
"b2": [
0.0
]
}
},
"G0_p2": {
"ok": true,
"note": "Rewards IDF coverage most, since it emphasizes discriminative disease terms, plus embedding, trigram and token overlap as supporting signals. A small length-ratio penalty in two hidden units discourages mismatched caption lengths. Weights are moderate, positive, and hand-set, not fitted to data.",
"cost_usd": 0.005186,
"cache_hit": true,
"params": {
"W1": [
[
1.5,
0.5,
0.3,
0.5,
0.0
],
[
0.8,
0.2,
0.8,
1.0,
0.0
],
[
1.2,
-0.3,
0.2,
0.3,
-0.3
],
[
-0.5,
0.6,
0.4,
-0.2,
-0.6
]
],
"b1": [
0.0,
0.0,
0.0,
0.2
],
"w2": [
1.0,
0.8,
0.6,
0.3
],
"b2": [
0.0
]
}
},
"G1_p0": {
"ok": true,
"note": "Rewards IDF coverage most, since it highlights discriminative terms shared with the test caption, plus Jaccard overlap, trigram and embedding similarity as secondary signals. A small penalty on length mismatch limits generic long captions. The weights are hand-set, not trained.",
"cost_usd": 0.011508000000000001,
"cache_hit": false,
"params": {
"W1": [
[
1.5,
0.5,
0.3,
0.5,
0.0
],
[
0.8,
1.2,
0.8,
0.0,
0.0
],
[
0.3,
0.0,
0.5,
1.5,
0.0
],
[
-0.5,
-0.3,
0.0,
0.0,
-1.0
]
],
"b1": [
0.1,
0.0,
0.0,
0.2
],
"w2": [
1.2,
0.8,
0.8,
-0.3
],
"b2": [
0.0
]
}
},
"G1_p1": {
"ok": true,
"note": "IDF coverage is the strongest signal for disease-specific terms, so it gets the largest weight. Embedding cosine and overlap/trigram add semantic and lexical agreement. A small penalty unit on caption length mismatch and low similarity demotes generic same-crop distractors.",
"cost_usd": 0.011764,
"cache_hit": false,
"params": {
"W1": [
[
1.5,
0.5,
0.3,
0.5,
0.0
],
[
0.4,
1.2,
0.8,
0.2,
-0.3
],
[
0.8,
-0.2,
0.2,
1.4,
0.0
],
[
-0.6,
-0.3,
0.0,
-0.4,
-0.6
]
],
"b1": [
0.0,
0.0,
0.0,
0.3
],
"w2": [
1.2,
0.6,
0.9,
-0.3
],
"b2": [
0.0
]
}
},
"G1_p2": {
"ok": true,
"note": "Rewards IDF coverage most, since disease-specific terms discriminate same-crop distractors. Jaccard, trigram and embedding similarity add smoother support. A mild length-ratio penalty in one unit discourages mismatched caption lengths. All weights are hand-set, not trained or validated.",
"cost_usd": 0.011723999999999998,
"cache_hit": false,
"params": {
"W1": [
[
1.5,
0.5,
0.3,
0.2,
0.0
],
[
0.4,
1.2,
1.0,
0.3,
-0.2
],
[
0.3,
0.2,
0.4,
1.5,
0.0
],
[
1.0,
-0.3,
0.0,
0.0,
-0.4
]
],
"b1": [
0.0,
0.0,
0.0,
0.1
],
"w2": [
1.0,
0.6,
0.7,
0.4
],
"b2": [
0.0
]
}
},
"G2_p0": {
"ok": true,
"note": "Hand-set from the dev pairs. Embedding cosine and trigram similarity most often separate same-disease references from same-crop distractors, so they get the largest weights. IDF coverage and overlap get smaller weights. Length mismatch is mildly penalized. One unit with a negative output weight penalizes low overall similarity.",
"cost_usd": 0.032714,
"cache_hit": false,
"params": {
"W1": [
[
0.2,
0.3,
0.5,
0.6,
-0.3
],
[
0.1,
0.2,
0.3,
0.8,
0.0
],
[
0.0,
0.2,
0.3,
0.3,
-0.2
],
[
-0.2,
-0.2,
-0.3,
-0.5,
0.1
]
],
"b1": [
0.0,
-0.5,
0.3,
0.0
],
"w2": [
0.8,
0.6,
0.3,
-0.5
],
"b2": [
0.0
]
}
},
"G2_p1": {
"ok": true,
"note": "Heuristic design, not fitted to the dev examples. Scores rise with IDF coverage and embedding cosine, with smaller weight on token and trigram overlap. A second unit rewards strong overall similarity. Two units penalize large caption length mismatch in either direction.",
"cost_usd": 0.030054,
"cache_hit": false,
"params": {
"W1": [
[
0.8,
0.2,
0.2,
0.6,
0.0
],
[
0.5,
0.3,
0.3,
0.3,
0.0
],
[
0.0,
0.0,
0.0,
0.0,
1.0
],
[
0.0,
0.0,
0.0,
0.0,
-1.0
]
],
"b1": [
0.0,
-0.5,
0.0,
0.0
],
"w2": [
0.6,
0.3,
-0.2,
-0.2
],
"b2": [
0.0
]
}
},
"G2_p2": {
"ok": true,
"note": "Hidden units 1 and 2 form a linear score from IDF coverage, trigram, overlap and embedding cosine, with IDF and trigram weighted highest. Units 3 and 4 penalize absolute length mismatch mildly. Hand-set from the development examples, not fitted.",
"cost_usd": 0.031042,
"cache_hit": false,
"params": {
"W1": [
[
0.8,
0.4,
0.6,
0.3,
0.0
],
[
-0.8,
-0.4,
-0.6,
-0.3,
0.0
],
[
0.0,
0.0,
0.0,
0.0,
1.0
],
[
0.0,
0.0,
0.0,
0.0,
-1.0
]
],
"b1": [
0.0,
0.0,
0.0,
0.0
],
"w2": [
1.0,
-1.0,
-0.2,
-0.2
],
"b2": [
0.0
]
}
},
"G3_p0": {
"ok": true,
"note": "Hand-set heuristic: positive weights on lexical overlap, trigram and embedding similarity, with IDF coverage weighted most. A mild penalty on length mismatch and a low-similarity unit limit weak matches. Not fitted or validated beyond a rough reading of the development examples.",
"cost_usd": 0.033676,
"cache_hit": false,
"params": {
"W1": [
[
1.0,
0.8,
0.5,
0.6,
0.0
],
[
0.5,
1.0,
1.0,
-0.3,
0.0
],
[
0.6,
0.3,
0.2,
1.2,
-0.3
],
[
-0.5,
-0.3,
-0.3,
-0.5,
0.0
]
],
"b1": [
0.0,
0.0,
0.0,
0.3
],
"w2": [
0.6,
0.5,
0.6,
-0.3
],
"b2": [
0.0
]
}
},
"G3_p1": {
"ok": true,
"note": "Hidden units 1 and 2 form a signed linear blend of IDF coverage, overlap, trigram and embedding cosine, weighted toward coverage. Unit 3 adds a bonus for high embedding similarity. Unit 4 mildly penalizes captions much longer than the reference. These weights are hand-set, not fitted.",
"cost_usd": 0.038432,
"cache_hit": false,
"params": {
"W1": [
[
0.8,
0.6,
0.4,
0.6,
0.0
],
[
-0.8,
-0.6,
-0.4,
-0.6,
0.0
],
[
0.0,
0.0,
0.0,
1.0,
0.0
],
[
0.0,
0.0,
0.0,
0.0,
1.0
]
],
"b1": [
0.0,
0.0,
-0.5,
-0.8
],
"w2": [
1.0,
-1.0,
0.4,
-0.3
],
"b2": [
0.0
]
}
},
"G3_p2": {
"ok": true,
"note": "Mostly monotone in lexical and embedding similarity: IDF coverage and overlap dominate, since same-disease pairs share specific terms. Trigram and embedding add smaller weight. A hidden unit contrasts coverage against overlap and trigram. A small length-mismatch penalty. Hand-set from dev examples, not fitted.",
"cost_usd": 0.03385,
"cache_hit": false,
"params": {
"W1": [
[
1.0,
0.8,
0.6,
0.3,
0.0
],
[
0.5,
0.5,
0.2,
0.8,
0.0
],
[
0.8,
-0.3,
-0.4,
0.0,
0.0
],
[
0.0,
0.0,
0.0,
0.0,
-0.3
]
],
"b1": [
0.0,
0.0,
0.0,
0.0
],
"w2": [
0.8,
0.5,
0.3,
-0.1
],
"b2": [
0.0
]
}
}
}
}