JevK5 training queue

The hosted queue supports pinned JevK5 4B weights as its starting model. New datasets are bound to a model identity when the processor validates them. The model identity and base revision are included in the hashed job manifest, miner assignment, customer job response and accepted release. Existing manifests that predate model identities retain the original Kev 0.8B contract.

The decision API provides hosted predictions separately. The automatic workflow connects approved assignment and verified activation of accepted adapters; export consent and miner approval remain required. The local and testnet fleet retains its existing Kev contract.

The hosted coordinator is https://training.zils.ai. Approved miners use that URL as their configuration’s coordinator value. Self-hosted deployments use their own coordinator URL.

Install and create the reference

Use Linux or WSL 2, Python 3.13, a BF16-capable NVIDIA GPU, its CUDA driver and the repository's pinned model/signing dependencies. The implementation was checked on an RTX 4090. Allow at least 25 GB free disk for models and packages; GPU requirements depend on input length. CPU and MPS execution are not supported by this JevK5 queue runtime. The coordinator API itself needs no GPU.

From a fresh clone:

git clone https://github.com/ooo-hq/zils.git
cd zils
uv venv --python 3.13 .venv-kev
uv pip install --python .venv-kev/bin/python --torch-backend=cu128 \
  -r requirements/model.txt -r requirements/rehearsal.txt
.venv-kev/bin/python -c "import torch; assert torch.cuda.is_available(); assert torch.cuda.is_bf16_supported()"
.venv-kev/bin/python -m zils.jevk5 reference --out models/jevk5-reference

The reference command downloads and verifies the pinned model, tokenizer and configuration. It creates a small model.json reference rather than duplicating the full base weights. The model is alibiserikbay/JevK5, revision c4f7fdb3aeab5582336406e78d3bef11bf98833d; the upstream runtime is pinned at f26426d16f59e8bbe1470e5b162cc89329e29b29 in requirements/model.txt.

For an already downloaded model, set ZILS_JEVK5_BASE_DIR to its directory and use --no-download. This override must pass the same weight and tokenizer checksums; it cannot select another model. Model processes run offline.

Configure the services

Complete the Supabase resources, customer Auth, HTTPS edge and protected server environment setup in Supabase training. Set ZILS_TRAINING_MODEL=jevk5-4b-v0.3 in the API and processor environments. Start the API with the existing python -m zils.coordinator serve command and the processor with:

.venv-kev/bin/python -m zils.coordinator process \
  --state .private/queue-processor --reference models/jevk5-reference --device cuda

Use a persistent service supervisor for deployment. The processor refuses to start if its reference does not match the configured active model. The read-only GET /v1/config response identifies that active model; it is configuration, not a miner-health or inference-availability signal.

On an approved miner, create the protected hotkey/wallet configuration described in the queue guide, download the same reference, and run:

.venv-kev/bin/python -m miner.queue \
  --config .private/queue-miner.json --state .private/queue-miner \
  --reference models/jevk5-reference --device cuda

The miner receives only approved training exports and signed storage URLs. It must not receive the processor's Supabase credential or calibration/test data. Run miner and processor under separate operating-system identities if they share a machine, with separate private state directories and environment files.

When multiple identities share one GPU, configure the same ZILS_COMPUTE_LOCK path for them. Pre-create it in an administrator-owned directory with group read/write access for the service identities; do not make the parent directory writable by miners. Without this override the existing per-user/device lock is used. Other applications need their own resource coordination.

Data and training contract

This version supports choice, true/false and ordinal-score decisions with at most 16 outcomes and 2,048 prompt tokens per example. Validation checks those limits before miner approval. Inputs are never silently truncated. The website's decision text is supplied as JevK5's criterion when the question has no explicit instructions; the original evidence remains present.

Training uses one epoch, attention-only rank-16 LoRA, alpha 32, dropout 0.05, learning rate 0.00002, batch size one and accumulation of four examples. Partial final batches use their actual size. The loss is cross entropy over the allowed answer-letter logits, not generated text. Base weights and the language-model head remain frozen. The candidate's raw temperature is 1.0; the validator fits the baseline and candidate separately using only calibration labels.

JevK5 candidates contain adapter_config.json, adapter_model.safetensors, and model.json; accepted downloads add release.json. Training keeps FP32 LoRA parameters and saves them in BF16, producing an approximately 29 MB adapter that fits a 50 MB storage upload limit. The loader also accepts earlier FP32 adapters. The validator instantiates its own fixed LoRA architecture and checks the exact tensor names, shapes, dtypes and finiteness before loading. It never executes miner-provided model code or loads a miner-chosen base. Artifact hashes include the model metadata and weights.

Calibration, positive uniform skill, the customer's accuracy target and strict Brier improvement remain required for an accepted trained model. A strong base may leave no qualifying fine-tuned candidate. no_qualifying_model remains a valid completion; the queue does not relabel the unmodified base as a trained customer model.

Existing jobs and rollback

Keep existing job rows, manifests, references and release files unchanged. The API determines old download filenames from each job's model, not today's active setting. To evaluate older pending jobs with the new processor, also provide --additional-reference models/reference and retain that Kev base cache. Miners must use a reference matching their assigned job; an incompatible miner refuses the assignment instead of training another model.

Before changing the active model, finish or inventory in-flight work and prepare the appropriate miners and references. Change the API environment and primary processor reference together. Rollback uses the same procedure with ZILS_TRAINING_MODEL=kev-0.8b-v1 and the Kev reference, retaining JevK5 as an additional reference if any JevK5 jobs still need evaluation. Do not change a prepared job's identity or rewrite its historical result.

Verification

make check covers both model formats through signed upload, training fixtures, calibration, acceptance and version-correct downloads, plus the original local and testnet contracts. Fixture tests establish workflow behavior, not model quality. Verify actual GPU training, artifact reload and held-out evaluation on authorized data before operational cutover. Neither a successful workflow nor a synthetic accuracy score establishes customer-task reliability.