maxdata-train: 15,257 records (E2-big teacher + v2.1 scripts + neg1 + neg1h), text only (K-34)
e2db892 verified | # maxdata-train: every training record we have for the 4B student (K-34) | |
| `maxdata_train.jsonl`: 15,257 records (format `mcp-v1`: `tools` + `messages`), built 2026-10-27 02:45 UTC with | |
| `tools/nav/mcp_remix.py maxdata --script script3_success_clean.jsonl` (repo KimperYang/VLMAuditor). Text only: the images are | |
| not in this folder (see "Images"). | |
| | part | records | C1 | C2 | C3 | G1 | G2 | G3 | S2 | T1 | T2 | T3 | V1 | V2 | V3 | V4 | | |
| |---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---| | |
| | teacher (E2-big) | 3,538 | 145 | 345 | 24 | 359 | 514 | 286 | 241 | 143 | 215 | 305 | 255 | 71 | 71 | 564 | | |
| | scripts v2.1 | 7,562 | 578 | 543 | 202 | 663 | 1157 | 465 | 423 | 494 | 836 | 193 | 1067 | 432 | 235 | 274 | | |
| | hard negatives neg1 | 1,630 | 5 | 0 | 8 | 230 | 666 | 115 | 117 | 25 | 10 | 0 | 195 | 35 | 134 | 90 | | |
| | hard negatives neg1h | 2,527 | 7 | 0 | 6 | 816 | 391 | 161 | 196 | 22 | 35 | 0 | 153 | 101 | 271 | 368 | | |
| | total | 15,257 | 735 | 888 | 240 | 2068 | 2728 | 1027 | 977 | 684 | 1096 | 498 | 1670 | 639 | 711 | 1296 | | |
| - **Teacher (3,538):** E2-big as it is: every clean teacher success (one report, the right one) of our runs: Gemini 3.8 Flash | |
| 1,703, GPT-5.6 846, GPT-6 / 6.1 413, Qwen3.8-27B 576. | |
| - **Scripts (7,562):** this dataset's `script3_success_clean.jsonl` (7,806 records, generator v2.1, thoughts written), minus | |
| the validation split (crc32(id) % 33 == 0: 244), as for every set since E1; deduplicated by uid (no duplicates found). | |
| - **Hard negatives (4,157, all):** a teacher's episode up to a wrong report, then an inspect of the evidence, one line saying | |
| the object looks normal and the teacher's next action (turn 1,817, move 695, done 1,113, look 510, wait 22). neg1 from | |
| katrina1's Gemini teachers, neg1h from the H100 teachers (`tools/worldgen/mcpgen_neg.py`). | |
| - No v1 scripts, no identification records. No record comes from the 7 held-out evaluation worlds. Hard classes | |
| (C1 C3 T1 T2 V1 V2 V3) are 37.9% of the records. | |
| - Records are our own procedurally built worlds; no benchmark tasks, rubrics or answers. The `llm` metadata of the records has | |
| no endpoint (removed for this public copy; training does not read it). | |
| ## Images | |
| Paths are relative to one IMAGE_ROOT. They are the union of: the IMAGE_ROOT of E2 (`e2-train/` of the private | |
| `kimperyang/vlmauditor-mcpgen-data`: its snapshots 20261024-0105, 20261025-0530, h100-20261025-1712), `allteacher-train/` and | |
| `neg1h/` there, katrina1's snapshots 20261026-0850, 1230, 1406, 1857 there, and this dataset's `images-script3-000..004.tar`. | |