d3-mini body: six LoRA runs averaged and scaled by 1.25 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.

Body

The body's sha256 is e666cbd24a6ee61abfe917ddfbb6fbdfce8d37992c942183ca0b6c7d7f20558c. It was built on 2026-11-12 at 08:42:59Z by our mean script mini_mean_v3.sh (sha256 prefix 757506e1). The body is the start plus 1.25 times the average change of the six runs:

body = start + 1.25 x (1/6) 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 six runs starts from MS11, using the flags in run_mini_seed_v5.sh (sha256 prefix dbd54e3c):

  • adapters: LoRA rank 32, alpha 64; no adapters on the MLP layers (MLP=0)
  • learning rate 5e-5, for 2 epochs
  • LoRA weight decay 1.0 (--lora-wd 1.0)
  • loss: cross-entropy (weight 1) plus Brier score (weight 1); KL weight 0
  • pin term: --pin-rows 532, --pin-weight 1.0, --pin-train 0.1

The runs differ only in their seed. These are the six run bodies:

Seed Body sha256
4402 814417fe3cfbd1e9fb9d9d42a6e0bf75105603114016e0d9df8ef295a5a40cbd
4412 f0d9d72dd2cdc9de6970e36eb8836d4a3e6ad79321bc8751fcdf6b4dccdfc156
4422 9dc31cd88eacc6e6c9a343af3799d1a632e5ab34b1fa34af2454e7ab6eed5909
4432 033b1b0f3bf383e25bd431aa10a28d2b889b49c5b2fff622f7fb8e14ae108fa9
4442 dc9003f23e8909845a31f0d752d6e46483512be39161fd1c01623e15ab7933c8
4452 3007c013aac063280f8ab4df1c562163ee88d7187423101555be4fc97d150c22

Data

The training file has 2,696 rows (sha256 prefix fb7552df). This is its own description:

all hard labels are writer answers; 2692 rows by team Claude (1685 worker3, 1007 worker2), 4 local Phi-4 arts-known-choice-k3 rows; no dataset gold

Each row also names the public training set whose row served as its topic or style example. NOTICE lists these sets with their counts.

Evaluation

This body has one read: the team's C3 read (read_body v2), compared with the MS11 start on 10 splits. The read file's sha256 prefix is d047f21a.

  • Mean change: +1.14. The splits range from -0.70 to +5.22.
  • Splits: 8 of 10 are above 0, and 2 of 10 are at +1.20 or more.
  • Our pre-set bar: +2.00. This read is below it.
  • Per-family check: fails. The retrieval-score family's mean change is -7.75, against a floor of -3.0.
  • Not read: the hold-out read, the copy check and the delta-correlation check.

This card reports no other results.

Licence

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

Downloads last month
-
Safetensors
Model size
5B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for tensor-tailor/zils-mini-body-31

Finetuned
Qwen/Qwen3.5-4B
Finetuned
(1)
this model