Zils lite reasoning v1

Experimental full BF16 body for subnet 127's d3-lite pool. Use the pinned d3-lite-0.8b-v1 chassis for its tokenizer, runtime and answer layer. Local proxy evaluation does not establish a private-exam score or incentive. See the pool contract for the submission requirements.

The starting model was the equal-weight Charlie320 v9a / qhoenix decision-v9 soup. A first 13,000-row distillation was followed by this one 12,000-row continuation: 9,000 fresh public TRAIN groups and 3,000 optimization-only replay rows. New targets mix 70% source labels/votes with 30% pinned Relay teacher output. Reasoning, language and arts receive 80% of the new rows. One third of eligible choice questions become candidate-correctness yes/no questions. All fresh training, retention, development and confirmation groups are disjoint from each other and excluded from our earlier data. Upstream source exposure is unknown.

The continuation used one RTX 4070 Ti SUPER with 16 GiB, LoRA rank 64, alpha 128, learning rate 5e-5, accumulation 16, seed 32 and 750 optimizer steps. The LoRA was merged into full body weights. The answer layer and vision tower stayed fixed. The first-stage body failed its own local submission gate; training.json preserves that unfavorable result, pinned lineage, source mixture and limitations.

Public proxy evaluation

Most development improvement is on yes/no questions generated from choice tasks. Training and evaluation share this candidate-correctness template. Choice-task skill also improves, by a much smaller amount. The large aggregate gain therefore measures transfer within this public format, rather than a comparable gain on the private validator exam. evaluation.json includes the type decomposition.

Scores follow the validator Brier/RPS formula with separate 20% calibration and present-cell weight renormalization. They cannot be compared directly with a nightly board score. evaluation.json includes all development diagnostics, five calibration splits, conditional paired intervals, stock comparisons and the once-used confirmation. The public confirmation samples original TRAIN source groups, rather than an independent official benchmark. Synthetic results measure transfer within a shared generator. Copy and rate checks are local estimates; private admission, rate, seating and payment remain unconfirmed.

Attribution

Apache-2.0. Original available upstream cards and notices are preserved under upstream/; see LICENSE and NOTICE. Source text is not redistributed. This is an experimental miner body, not a released Zils model or a standalone general text-generation model. Full preparation scripts and transformed data are not bundled, so this is a methodology report rather than a reproducible release.