zils-mini-body-35: one full fine-tune of the MS11 start on our lh4000 data (every language weight except the tied embedding)

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.

Our own simple test reads this body below the MS11 start (see Evaluation). We submit it to learn its real score.

Body

The body's model.safetensors sha256 is 412eb3e2946885af4e6560a32b4dc69d9a19953e816da6baba34bd048c209a4b; its Zils body digest is f5e332bb3df24e6b8eaa488c42e49033ac7948c3a2d09d49b0a39d2697aa6163. It is the final weights of one training run; it is not a mean of runs and it is not scaled.

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

One run starts from MS11 and uses our trainer train_full_gold_v4m2.py (sha256 prefix c0f98250), seed 5101. These values are read back from the run's train_meta.json, GOLD_ONLY_RECEIPT.json and log; the fp32 master weights, the frozen vision tower and weight decay 0 are fixed in the trainer's code:

• full fine-tune, no adapters: 425 language-model tensors (3,570.1M parameters) are trained as fp32 master weights under bf16 autocast
• frozen: the token embedding (tied to the output head) and the vision tower
• optimizer: 8-bit AdamW from bitsandbytes 0.49.0; learning rate 5e-6, weight decay 0, 50 warm-up steps
• 1 epoch over 3,982 rows (0 skipped); batch 8 with 2 accumulation steps (249 steps)
• loss on every row: soft cross-entropy on the gold label (weight 1) plus Brier score (weight 1) plus KL to the frozen MS11 start (weight 0.1); no pin rows; label smoothing 0; RPS 0
• start check before training (stored start distributions against a live start forward, 64 rows, two batch groupings): largest difference 0.0154 / 0.0181, mean 0.0013 / 0.0015; the trainer's gate is max 0.20, mean 0.01, p90 0.02

Data

The training file has 3,982 rows (sha256 prefix ef8d453b, FEED_lite_s10_LH4000_20261113T192002Z). Where the labels come from, counted from each row's label_method:

• 3,714 rows: the writer's own answer (2,066 by one Claude session, 1,372 by a second, 271 by a third, 4 by a local Phi-4 model, 1 by a local Mistral-Nemo model)
• 217 rows: the human gold label of a public TRAIN row (MNLI 156, QNLI 51, PIQA 8, SNLI 2)
• 15 rows: derived by code from the human answer spans of SQuAD v2
• 3 rows: the author's own label of a public TRAIN tweet (iSarcasmEval task A, its 'sarcastic' column)
• 33 rows: computed by code from the item's own specification, with an independent recount

Most rows also name the public training set whose row served as their style example, or as their topic and style example (273 rows record none). NOTICE lists these sets with their counts.

Evaluation

These are our own reads, not an official score.

• Simple test (our read of 1,488 rows: the probability of the gold answer for each row, paired with MS11's own read of the same rows; the change in mean log p(gold) with a 95% bootstrap interval, 2,000 resamples):

  set	rows	p(gold), body vs MS11	accuracy, body vs MS11	change in mean log p(gold) [95%]	read
  our MINI hold-out rows	149	0.880 vs 0.903	0.919 vs 0.980	-0.084 [-0.189, +0.010]	not different
  C3 base (dev_gold + pool_known)	1,339	0.843 vs 0.862	0.895 vs 0.926	-0.064 [-0.092, -0.034]	below
  dev_gold	787	0.835 vs 0.877	0.886 vs 0.935	-0.118 [-0.153, -0.089]	below
  pool_known	552	0.855 vs 0.840	0.909 vs 0.913	+0.013 [-0.037, +0.060]	not different

  The test's rule: a body is "not below MS11" when neither the hold-out rows nor the C3 base is below. Here the C3 base is below, so this body reads below MS11.

• Copy screen (our sketch screen against our earlier bodies, with the current soup kept): no copy found in this screen. The largest share explained by an earlier body is 0.0012, by zils-mini-body-32 (our line 0.30). The behaviour distances are 0.0495 bits to MS11, 0.0486 bits to zils-mini-body-31, 0.0460 bits to zils-mini-body-32, 0.0441 bits to zils-mini-body-34 (our line: at least 0.01).

This card reports no other results.

Licence

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