Zils mini champion relay v1

An experimental full BF16 body for the Zils subnet 127 d3-mini pool, trained from joamCriff123/zils-mini-sx3. The starting body was the only member shown in the 2026-11-09 mini soup, with an unpaid seat. This submission's local results establish neither a paid seat nor performance on the private validator exam.

Use the exact d3-mini-4.5b-v1 chassis: runtime, prompt, tokenizer and answer layer are fixed. This repository supplies body weights, not a standalone generic text-generation model. The answer layer and vision tower were frozen. Full lineage, source revisions and settings are recorded in training.json. The submitted body is the merged trained candidate, without a post-training weight average.

Runtime and pool contracts are documented in the operator guide. The pinned chassis is zils-ai/d3-mini-4.5b-v1.

Training

Merged LoRA rank 64, alpha 128, dropout 0, learning rate 5e-5, AdamW without weight decay, accumulation 16, seed 24, 10,600 training rows / 663 optimizer steps, on one RTX A6000. Nine thousand frozen public Relay TRAIN rows use 40% pinned27B teacher probabilities, 30% original gold distributions and 30% starting mini probabilities on those same inputs. This one corrective run adds same-input retention and more gold support after the initial run lost seed and Outcomes score. Another 1,600 separate public TRAIN rows retain the starting mini model's own probabilities. A further 1,000 retention rows were held aside and not optimized. Training did not use seed or evaluation rows. Upstream exposure to these public sources is otherwise unknown.

Public proxy evaluation

Fresh confirmation difference: +3.316 points, conditional 95% interval [+1.926, +4.999]. See evaluation.json for all evaluated development variants, actual uniform weight-average soup, calibration sensitivity, exclusions, bootstrap bounds and copy/rate diagnostics. Unfavorable results are retained there.

On the same confirmation draw, pinned stock mini scored 64.379 against the submitted body's 62.814: candidate minus stock -1.565 points (conditional 95% interval [-4.525, +0.597]). This proxy does not establish a win over the best house model. Local fine-tune base diagnostics concern the measured contributing starting soup; a paid seat still requires useful private-round soup contribution.

The starting model explicitly trained on public seed exams v1 and v4, so seed scores are diagnostic and were excluded from the primary decision criterion. Development rows and the relay recipe had been used in an earlier lite experiment. The 1,500-row confirmation draw newly excludes every earlier lite train/dev/confirmation source group as well as this experiment's training and development groups. These held groups are from public original TRAIN splits; they are not a sealed or official independent test set.

The actual validator uses different items and may reject, decline to seat or decline to pay this body. A fine-tune's private score must clear its measured base's performance gate; local nominal trainer/scout classification alone does not establish payment. No mining revenue or release claim is made here.

An additional actual 50:50 weight average of pinned stock mini and the fixed submitted body scored 64.626 on the same confirmation rows, versus stock 64.379: +0.247 points (conditional 95% interval [-1.952, +1.397]). This diagnostic was added after observing the submitted body's confirmation score below stock; it reuses exposed rows and is not an independent confirmation or a submission-candidate selection. The +0.247-point blend improvement is below the protocol's 0.5 greedy-addition margin and 0.8 paid-seat contribution floor on this proxy. It does not demonstrate paid-seat readiness. The submitted weights remain the original fixed relay24 body. See the full additional diagnostic in evaluation.json.

Attribution

Apache-2.0. Upstream card and notices are preserved in upstream/; see LICENSE and NOTICE. The starting model reports soup lineage from qhoenix/zils-mini-decision-v2 and Charlie320/zils-mini-s1, plus zils-ai/d3-mini-4.5b-v1 and Qwen/Qwen3.5-4B. Public training text retains its original source licences. No source text or held-out examples are redistributed.