zils-mini-body-34: three LoRA runs on our lh3500 v2 data, averaged and scaled by 1.5 on the MS11 start

This is a full d3-mini body for the Zils d3-mini pool. It uses the model d3-mini-4.5b-v1 from the Zils repository at revision e9e87b3e.

This body is below our own pre-set bars (see Evaluation). We submit it to learn its real score.

Body

The body's model.safetensors sha256 is 4d29232e55d2ef504115f1ce5468bf0913c2d8f006de4c77a46ac7485a5e29ff; its Zils body digest is bb98c77d7539ee2be41551f86bb535f1bcb7d1ad97a8558749f6a29a48b88e79. It was built by our combination tool delta_combo.py v1.4 (sha256 prefix 0a9b49b1; build receipt sha256 prefix 97e457c5) as the start plus 1.5 times the average change of the three runs, on every tensor (152 tensors differ from the start):

body = start + 1.5 x (1/3) x sum(run - start)

Start (MS11)

MS11 is the official soup of the d3-mini round of 2026-11-11 (soup hash 9a21f188dc04284d8ef526870e8a37b19533556eabaccc31aecf1457b0b3ce1a). It was rebuilt with the Zils repository at revision e9e87b3e from the soup's two public member bodies:

• qhoenix/zils-mini-decision-l13_avg678 at revision 091c6ebe (body sha256 prefix 11bfca0b)
• joamCriff123/zils-mini-sx3 at revision c9432ada (body sha256 prefix e0ccf728)

The start body's sha256 is 8c26d1ce07e3d26e6e9b04c3cdff13dafc37fbcab70d7c18e40d14cb78fdd57c. The build receipt's sha256 prefix is 02af29da.

Training

Each of the three runs starts from MS11 and uses our trainer train_soft.py. These values are read back from each run's train_meta.json:

• adapters: LoRA rank 32, alpha 64, dropout 0.05; no adapters on the MLP layers
• learning rate 5e-5, for 2 epochs; batch 8 with 2 accumulation steps (320 steps)
• LoRA weight decay 1.0
• loss: cross-entropy (weight 1) plus Brier score (weight 1); KL weight 0
• pin term: pin weight 1.0, --pin-train 0.1 with the live pin; the pin file is empty (0 pin rows)

The runs differ only in their seed. These are the three run files, as the build receipt and the run logs name them:

Seed	Run model.safetensors sha256	Run body digest
4762	96cb1facaefee281f3a8dc873b8154c2a8c394f62b37f64f062c94213a464a5c	1d885ee6f7b6837b85f0fd437fede97d813fa0d9412678804a24d021b5ea5da3
4772	2c9f53a879fb6b8fb0c21ae7e7182aacc33e27aca7099be20ad3efaa08e940c7	3bcffecf1eb5d8634e37ecee3db90d1e0ab92fc42745ca67c775d3fb1ec34711
4782	4c5da3fffc883b70a5c9620a49b876b27fd83187fb265ad41e5228019fd19651	d52e67a8d73f2e4d57ce9e7de106796e1c29b32e04294e8726ba684108666c67

Data

The training file has 2,560 rows (sha256 prefix 6feec371). This is its own description, as each run records it:

lh3500v2 feed FEED_lite_s10_LH3500V2_20261113T015224Z 6feec371 (worker1 seal 02:06Z, result #4098; builder lite5000_feed_v24.py f319ab30)

Where the labels come from, counted from each row's label_method:

• 2,390 rows: the writer's own answer (1,198 by one Claude session, 1,190 by another, 2 by a local Phi-4 model)
• 145 rows: the human gold label of a public TRAIN row (MNLI 105, QNLI 38, SNLI 2)
• 20 rows: derived by code from the human answer spans of SQuAD v2
• 4 rows: the author's own label of a public TRAIN tweet (iSarcasmEval task A, its 'sarcastic' column)
• 1 row: computed by code from the item's own specification

Most rows also name the public training set whose row served as their style example, or as their topic and style example (172 rows record none). NOTICE lists these sets with their counts.

Evaluation

These are our own reads, not an official score.

• C3 read (read_body v2): this body against the MS11 start, on the same items, in points, on 10 registered splits.

  • Mean change: -0.38. The splits range from -2.45 to +3.04.
  • 2 of 10 splits are above 0, and 2 of 10 are at +2.00 or more.

  category	mean change
  reasoning	-0.80
  language	+1.52
  retrieval	-3.87
  tools	+0.96
  arts	+0.32

• C3 with category weights (reasoning 0.26, language 0.26, retrieval 0.20, tools 0.18, arts 0.10): -0.75, with 2 of 10 splits above 0. Our own line for this read is +2.8, so the read is below it.

• Hold-out read (mean change in the probability of the gold answer, our sealed hold-out rows): +0.0515 on the 140 exam-length rows (+0.0514 on all 149). This passes our hold-out line.

• Per-family check: fails. 3 families are below the floor of -3.0: retrieval-score at -10.38, language yes/no at -5.82, tools-score at -3.05.

• Dev read (our dev set, 3,279 items): accuracy 78.59 against 79.54 for the start, a change of -0.94 points (standard error 0.40).

• Copy screen (our sketch screen against our earlier bodies, with the current soup): not a copy. The largest share explained by an earlier body is 0.249, by zils-mini-body-32 (our line 0.30). The behaviour distance is 0.0139 bits to zils-mini-body-32 and 0.0220 bits to MS11 (our line: at least 0.01).

• Our commit bar. Our pre-set bar is a 10-split mean gain of at least +2.00 points over the start, with at least 8 of the 10 splits at +2.00 or more. This body is below our own pre-set bar (mean -0.38, 2 of 10 splits at +2.00 or more), and it fails our per-family check.

This card reports no other results.

Licence

The licence is the start's LICENSE, unchanged. NOTICE lists the sources this body depends on.