Upload prepared BiCA SFT data

.gitattributes

@@ -58,3 +58,12 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+sft_bica/eval.jsonl filter=lfs diff=lfs merge=lfs -text
+sft_bica/train_bica.jsonl filter=lfs diff=lfs merge=lfs -text
+sft_bica/train_fever.jsonl filter=lfs diff=lfs merge=lfs -text
+sft_bica/train_fiqa.jsonl filter=lfs diff=lfs merge=lfs -text
+sft_bica/train_hotpotqa.jsonl filter=lfs diff=lfs merge=lfs -text
+sft_bica/train_msmarco.jsonl filter=lfs diff=lfs merge=lfs -text
+sft_bica/train_nq.jsonl filter=lfs diff=lfs merge=lfs -text
+sft_bica/train_squadv2.jsonl filter=lfs diff=lfs merge=lfs -text
+sft_bica/train_trivia.jsonl filter=lfs diff=lfs merge=lfs -text

sft_bica/bica_meta.json

@@ -0,0 +1,16 @@
+{
+  "source": "bisectgroup/hard-negatives-traversal (BiCA, arXiv 2511.08029)",
+  "raw_rows": 19997,
+  "usable_rows": 12069,
+  "written_rows": 72885,
+  "upsample": 6.039,
+  "negatives_in": 129971,
+  "negatives_vetoed": 482,
+  "negative_fn_rate": 0.0037,
+  "rows_dropped_short": 7928,
+  "num_negatives": 7,
+  "judge": "lightonai/GTE-ModernColBERT-v1",
+  "veto": "negative >= pair's own positive (per-negative)",
+  "target_formula": "usable^0.5 * (1 - fn_risk), calibrated on fiqa",
+  "seed": 42
+}

sft_bica/eval.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc3dbfdfbd7ff16c593365d4768e1f6fa762d041e908bea28804a3260f540ea8
+size 11394235

sft_bica/prepare_meta.json

@@ -0,0 +1,100 @@
+{
+  "recipe": "judge-vetoed band sampler + quality-weighted mixture (notes 19.5-19.6) + BiCA citation-traversal biomedical source (8th)",
+  "judge": "lightonai/GTE-ModernColBERT-v1",
+  "veto": "cand >= pair's own positive",
+  "band_lo": 0.9,
+  "band_hi": 0.98,
+  "num_negatives": 7,
+  "budget": 1500000,
+  "seed": 42,
+  "mixture": "target ~ usable^0.5 * (1 - fn_risk)",
+  "risk": {
+    "fiqa": 0.2075,
+    "squadv2": 0.0277,
+    "hotpotqa": 0.0106,
+    "fever": 0.0281,
+    "msmarco": 0.2404,
+    "nq": 0.3263,
+    "trivia": 0.0841
+  },
+  "sampling": "random 7 within band per emitted row (diverse across upsampled copies)",
+  "targets": {
+    "fiqa": 31979,
+    "squadv2": 218202,
+    "hotpotqa": 235441,
+    "fever": 227957,
+    "msmarco": 338064,
+    "nq": 150738,
+    "trivia": 297619
+  },
+  "per_split": {
+    "fiqa": {
+      "usable": 3672,
+      "target": 31979,
+      "written": 31900,
+      "skipped": 0,
+      "upsample": 8.71
+    },
+    "squadv2": {
+      "usable": 113579,
+      "target": 218202,
+      "written": 217595,
+      "skipped": 0,
+      "upsample": 1.92
+    },
+    "hotpotqa": {
+      "usable": 127703,
+      "target": 235441,
+      "written": 234817,
+      "skipped": 0,
+      "upsample": 1.84
+    },
+    "fever": {
+      "usable": 124063,
+      "target": 227957,
+      "written": 227340,
+      "skipped": 0,
+      "upsample": 1.84
+    },
+    "msmarco": {
+      "usable": 446692,
+      "target": 338064,
+      "written": 337991,
+      "skipped": 0,
+      "upsample": 0.76
+    },
+    "nq": {
+      "usable": 112900,
+      "target": 150738,
+      "written": 150738,
+      "skipped": 0,
+      "upsample": 1.34
+    },
+    "trivia": {
+      "usable": 238126,
+      "target": 297619,
+      "written": 297619,
+      "skipped": 0,
+      "upsample": 1.25
+    }
+  },
+  "eval_rows": 2000,
+  "multi_source": true,
+  "added_source": {
+    "source": "bisectgroup/hard-negatives-traversal (BiCA, arXiv 2511.08029)",
+    "raw_rows": 19997,
+    "usable_rows": 12069,
+    "written_rows": 72885,
+    "upsample": 6.039,
+    "negatives_in": 129971,
+    "negatives_vetoed": 482,
+    "negative_fn_rate": 0.0037,
+    "rows_dropped_short": 7928,
+    "num_negatives": 7,
+    "judge": "lightonai/GTE-ModernColBERT-v1",
+    "veto": "negative >= pair's own positive (per-negative)",
+    "target_formula": "usable^0.5 * (1 - fn_risk), calibrated on fiqa",
+    "seed": 42
+  },
+  "note": "Identical to data_judged2048_soft for the 7 original sources (symlinked, byte-identical). The only variable vs sft_soft_gb512_lr12 is the presence of train_bica.jsonl. BiCA negatives were vetoed per-negative in GTE-ModernColBERT MaxSim space, the same rule and space as the other 7; measured per-negative FN rate 0.50%, row rate 1.73% -- second cleanest source after hotpotqa."
+}

sft_bica/train_bica.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dd355662f1e8a572d24a8e0604e9710a4d8e2ab253e61f87d118aeb4a7e14131
+size 816036442

sft_bica/train_fever.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:432d1b217abc04808490002bd328baa127b97f38f00ac132383bd743adb45898
+size 1663040430

sft_bica/train_fiqa.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:daae6841d082c5653f71a66015d420b35b8285405d3b645e3c9c0b442935fd6d
+size 244747182

sft_bica/train_hotpotqa.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2c14d2f5877f2d29ebafcf3425b159164e12cda1b61f47e99c09867391320f8c
+size 713361106

sft_bica/train_msmarco.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d190d22109ef885984eb19dc2a05cc5f25bbac06ef710ad9ecde14c9d6db8400
+size 1007742459

sft_bica/train_nq.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ec871966038a10cfe54d944dd1ef2fb8a7553595288187d921490263f1e7faf9
+size 881312702

sft_bica/train_squadv2.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4f8647b1cda9bcae929533ff3434d7ea33b3b52bc62e2d5621f9064c529d76d0
+size 1474211359

sft_bica/train_trivia.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7d78bb367707e61206a795b9aa9edc368f7ddd60385b79727b4739345ae70f6e
+size 1557091570