d3-lite body: jiwo + distilled LoRA student, 50/50 soup
A body for the Zils d3-lite soup pool (chassis d3-lite-0.8b-v1).
How it was made
- Start point.
zils-ai/jiwo-0.8b-full@adce2ce92cd2130fd3ed451cf5f9642e3dfe4a57(apache-2.0), built as a one-member body on the pinned chassisvllm-sr/d3-lite@03560c26f5199c3c04129a2ccb38960c913635e0(apache-2.0). - Teacher. The mini house soup
torchcast-ai/torchcast-decision-4b@492bfb19662f3c6e601cc9c4bd2911ac95609ed4plusvllm-sr/d3-mini@df99bf4ca4418bc6908ee68e9b7479af1f3b8800(both apache-2.0), built on the pinned mini chassis. - Student.
zils.teacher train --lora 64 --start <jiwo body> --seed 3: one pass of KL distillation on the teacher's answers, blended 0.7 teacher with 0.3 true label (21,470 rows, sha256d176e088308c93dd2ec88167ef9b4f072caae78527714485015883a13b24abf1). - Soup.
zils.teacher soup: 0.5 jiwo body + 0.5 student.
Training data
Items came only from sources whose collector query is keyed by the event's creation time:
- arXiv subject classification (
arxiv-category-*, keyed by submission date); - Stack Exchange closed and accepted outcomes (
stackexchange-closed-*,stackexchange-accepted-*, keyed by creation date).
Every item was asked before 2026-08-31 and at least 110 days before collection. The train split covers asked times 2026-05-19 to 2026-07-17 (51,137 items; 11,235 used). Articles for deletion, Gerrit and GH Archive were excluded from training, because their collectors' query day does not bound when an item was asked.
The standard nightly Zils Outcomes collection does not reach events this old. A validator that deliberately backfilled old days could still overlap with these items. We accept that as a residual risk; we cannot rule it out. We did not track, predict or target any validator round or backfill, and we will stop and report if overlap is found.
Provenance (sha256)
| Source | Revision | File or digest |
|---|---|---|
| Qwen/Qwen3.5-0.8B (parent) | 2fc06364715b967f1860aea9cf38778875588b17 | weights file 04b1c301231dd422b8860db31311ab2721511346a32cb1e079c4c4e5f1fe4696 |
| vllm-sr/d3-lite (chassis) | 03560c26f5199c3c04129a2ccb38960c913635e0 | weights file ca6222dbf2b72899ae45580ed84d26d1f0e69a8de212edda295072000e34a4c2; readout fe11cfac5c79b6a83e35b707581b1260214279bb22b003c215c62719186ef629 |
| zils-ai/jiwo-0.8b-full (start point) | adce2ce92cd2130fd3ed451cf5f9642e3dfe4a57 | weights file 4b0188ef1199c14027e8f2a08277d829ea17d8b8cf5078f85848120c7e5f795a; zils body digest c2321f4492c38e67d0c2e2c4d0d3b105ccefd526b6643e8748152f204446f5e2 |
| torchcast-ai/torchcast-decision-4b (teacher) | 492bfb19662f3c6e601cc9c4bd2911ac95609ed4 | zils body digest 0aba18eb6ef7277a9ca6f0c47103b87f147fbdd1f1d8ce84dbd329bfd6ae6d48 |
| vllm-sr/d3-mini (teacher) | df99bf4ca4418bc6908ee68e9b7479af1f3b8800 | zils body digest 34a7a0ac1bccb74547ad0b5d16230a4b465e8087dd114f8564fdd19b36c5c763 |
| teacher answers (blended) | d176e088308c93dd2ec88167ef9b4f072caae78527714485015883a13b24abf1 |
Data manifest
The original manifest has sha256 a0d8ea768af04100b5f59a1fe3168d9db8301c4e8569a8d4a8c08234a636cfdb and was generated with zils commit 5568921ac90e81abddc726e3e905eea5e20824b0, plan sha256 f7369566df932743e3e82926809ea68f80d94cea6c998efbf3897300d1000439. The copy below omits one internal field, policy.ruling, an approval note; nothing else is changed. The copy's sha256 is 905600eb6444560b6b2283c20ee83a2dbc35329effd6129b8d8c48021646f9d4.
{
"created_at": "2026-11-05T16:34:38.720725+00:00",
"dropped": {
"balanced:refused-family": 424,
"balanced:too-recent": 209,
"unbalanced:refused-family": 1973,
"unbalanced:too-recent": 849
},
"plan_sha256": "f7369566df932743e3e82926809ea68f80d94cea6c998efbf3897300d1000439",
"policy": {
"accepted_families": {
"arxiv-category-": "arxiv",
"stackexchange-accepted-": "stackexchange",
"stackexchange-closed-": "stackexchange"
},
"asked_before": "2026-08-31T00:00:00+00:00",
"blocked_licences": [
"unreviewed"
],
"cutoff": "2026-07-18T16:34:38.720725+00:00",
"min_age_days": 110,
"refused_sources": [
"afd",
"gerrit",
"gharchive"
]
},
"raw_sha256": "223ddd3c3baecce578a8216440c5df14c384c8f96c7f8abe6de24fda02c95b72",
"raw_unbalanced_sha256": "7c0c2f0fe48f8301efc76ae2dbd220b7e1541b6ae9f9252b0431739ef6f50711",
"splits": {
"dev": {
"asked_max": "2026-07-18T09:11:04+00:00",
"asked_min": "2026-05-21T00:00:03+00:00",
"families": {
"arxiv-category-cs": 5644,
"arxiv-category-math": 1841,
"arxiv-category-other": 955,
"arxiv-category-physics": 3247,
"stackexchange-accepted-arts": 5,
"stackexchange-accepted-language": 6,
"stackexchange-accepted-reasoning": 4,
"stackexchange-accepted-retrieval": 3,
"stackexchange-accepted-tools": 3,
"stackexchange-closed-arts": 9,
"stackexchange-closed-language": 6,
"stackexchange-closed-reasoning": 17,
"stackexchange-closed-retrieval": 9,
"stackexchange-closed-tools": 13
},
"items": 11762,
"labels": {
"arxiv-category-cs=A": 1434,
"arxiv-category-cs=B": 1410,
"arxiv-category-cs=C": 1424,
"arxiv-category-cs=D": 1376,
"arxiv-category-math=A": 495,
"arxiv-category-math=B": 454,
"arxiv-category-math=C": 452,
"arxiv-category-math=D": 440,
"arxiv-category-other=A": 267,
"arxiv-category-other=B": 230,
"arxiv-category-other=C": 224,
"arxiv-category-other=D": 234,
"arxiv-category-physics=A": 793,
"arxiv-category-physics=B": 806,
"arxiv-category-physics=C": 829,
"arxiv-category-physics=D": 819,
"stackexchange-accepted-arts=A": 1,
"stackexchange-accepted-arts=B": 3,
"stackexchange-accepted-arts=C": 1,
"stackexchange-accepted-language=A": 2,
"stackexchange-accepted-language=B": 3,
"stackexchange-accepted-language=C": 1,
"stackexchange-accepted-reasoning=A": 1,
"stackexchange-accepted-reasoning=B": 3,
"stackexchange-accepted-retrieval=A": 1,
"stackexchange-accepted-retrieval=B": 2,
"stackexchange-accepted-tools=A": 1,
"stackexchange-accepted-tools=B": 2,
"stackexchange-closed-arts=false": 6,
"stackexchange-closed-arts=true": 3,
"stackexchange-closed-language=false": 2,
"stackexchange-closed-language=true": 4,
"stackexchange-closed-reasoning=false": 10,
"stackexchange-closed-reasoning=true": 7,
"stackexchange-closed-retrieval=false": 5,
"stackexchange-closed-retrieval=true": 4,
"stackexchange-closed-tools=false": 4,
"stackexchange-closed-tools=true": 9
},
"licences": {
"CC-BY-SA-4.0": 75,
"CC0-1.0": 11687
},
"sha256": "399cd23dcf58780b93a63e7f9edf51fc1b4bdf0eb918077c5ec8e616a8563281"
},
"train": {
"asked_max": "2026-07-17T23:58:02+00:00",
"asked_min": "2026-05-19T00:00:19+00:00",
"families": {
"arxiv-category-cs": 23958,
"arxiv-category-math": 7827,
"arxiv-category-other": 4075,
"arxiv-category-physics": 14042,
"stackexchange-accepted-arts": 12,
"stackexchange-accepted-language": 18,
"stackexchange-accepted-reasoning": 45,
"stackexchange-accepted-retrieval": 13,
"stackexchange-accepted-tools": 24,
"stackexchange-closed-arts": 145,
"stackexchange-closed-language": 137,
"stackexchange-closed-reasoning": 363,
"stackexchange-closed-retrieval": 98,
"stackexchange-closed-tools": 380
},
"items": 51137,
"labels": {
"arxiv-category-cs=A": 6031,
"arxiv-category-cs=B": 5994,
"arxiv-category-cs=C": 5946,
"arxiv-category-cs=D": 5987,
"arxiv-category-math=A": 1888,
"arxiv-category-math=B": 1944,
"arxiv-category-math=C": 1933,
"arxiv-category-math=D": 2062,
"arxiv-category-other=A": 1067,
"arxiv-category-other=B": 1021,
"arxiv-category-other=C": 1008,
"arxiv-category-other=D": 979,
"arxiv-category-physics=A": 3557,
"arxiv-category-physics=B": 3455,
"arxiv-category-physics=C": 3524,
"arxiv-category-physics=D": 3506,
"stackexchange-accepted-arts=A": 3,
"stackexchange-accepted-arts=B": 3,
"stackexchange-accepted-arts=C": 4,
"stackexchange-accepted-arts=D": 2,
"stackexchange-accepted-language=A": 8,
"stackexchange-accepted-language=B": 8,
"stackexchange-accepted-language=C": 2,
"stackexchange-accepted-reasoning=A": 21,
"stackexchange-accepted-reasoning=B": 21,
"stackexchange-accepted-reasoning=C": 3,
"stackexchange-accepted-retrieval=A": 3,
"stackexchange-accepted-retrieval=B": 8,
"stackexchange-accepted-retrieval=C": 1,
"stackexchange-accepted-retrieval=D": 1,
"stackexchange-accepted-tools=A": 16,
"stackexchange-accepted-tools=B": 7,
"stackexchange-accepted-tools=C": 1,
"stackexchange-closed-arts=false": 128,
"stackexchange-closed-arts=true": 17,
"stackexchange-closed-language=false": 131,
"stackexchange-closed-language=true": 6,
"stackexchange-closed-reasoning=false": 326,
"stackexchange-closed-reasoning=true": 37,
"stackexchange-closed-retrieval=false": 87,
"stackexchange-closed-retrieval=true": 11,
"stackexchange-closed-tools=false": 350,
"stackexchange-closed-tools=true": 30
},
"licences": {
"CC-BY-SA-4.0": 1235,
"CC0-1.0": 49902
},
"sha256": "ce85e8e43b67fe1fd47c4b50049b781be1df57304cc4b55eac7d6ad31562c328"
}
},
"train_export": "unbalanced",
"zils_commit": "5568921ac90e81abddc726e3e905eea5e20824b0"
}
Licence and attribution
Apache-2.0, as are all the source models. Training data: arXiv metadata (thank you to arXiv for use of its open access interoperability) and Stack Exchange (CC-BY-SA-4.0). No Stack Exchange text is republished in this repository.