Pinned H2O text adapters
New text training jobs default to h2o-lightning-4b-v1.2.3:
| Component | Pin |
|---|---|
| Base | h2oai/h2o-lightning-4b |
| Revision | acaf0d4ea251e54de928c75ef4352670d33192d3 (v1.2.3) |
| Architecture | Qwen3_5ForConditionalGeneration, text path only |
| GPU runtime | Python 3.12, PyTorch 2.10.0 CUDA 13.0, Transformers 5.17.0, PEFT 0.21.0 |
| Kernels | FLA / fla-core 0.5.2, causal-conv1d 1.7.0, SDPA |
| Precision | BF16 backbone and serialized adapter, FP32 answer projection and live LoRA parameters |
The miner setup creates the isolated runtime. The coordinator and queue client retain Python 3.13. Set ZILS_H2O_RUNTIME_PYTHON for both miners and validators; neither silently substitutes the older JevK5 runtime. Use an NVIDIA driver compatible with CUDA 13.0. The runtime rejects missing fast kernels and different pinned package versions before loading a model.
zils/h2o-pins.json records file checksums, including the upstream prompt code and serving configuration. The runtime loads only this verified code from the operator's base cache. Miner artifacts cannot supply executable code or choose the base model. To reuse a verified download, set ZILS_H2O_BASE_DIR to its directory. Every required file is still checked.
Training and scoring
The supported text contract is choice, true/false, or score with up to 255 outcomes and 2,048 full prompt tokens. Oversized inputs and images are rejected; examples and options are never silently truncated. Image jobs retain ImaJev.
Prompts use H2O's native non-thinking chat template and Answer: prefix. Native temperatures are 0.75 for choice, 0.8 for true/false, and 0.65 for score; the native true/false probability floor is 0.801. These settings ignore ambient SHIM_* overrides. The validator fits a further scalar temperature to that native distribution using only the job's held-out calibration split. Its stored temperature=1.0 means no additional calibration, not replacement of native temperatures. True/false answers retain H2O's true-first internal option order.
Training reuses the queue's one-epoch AdamW recipe: learning rate 0.00002, batch one, accumulation four, rank-16 LoRA, alpha 32, dropout 0.05. Only the 152 language attention/gated-delta projections are adapted. Vision weights, embeddings and the output head stay frozen. Cross entropy uses the native type temperature. Partial final accumulation windows retain their actual size.
Candidates contain adapter_config.json, adapter_model.safetensors, and model.json. Metadata binds the entire H2O profile; loaders construct a fixed architecture and verify tensor names, shapes, dtypes and finiteness. A qualifying adapter must still improve held-out Brier loss and meet the customer's accuracy target. Choosing H2O does not guarantee that fine-tuning improves every task.
An RTX 4090 check on 2026-10-10 trained six synthetic examples covering all three question types at 2,040–2,048 prompt tokens. Two optimizer steps took 27.56 seconds and peaked at 9.53 GiB of allocated GPU tensor memory. All 304 language-only LoRA tensors were finite, and the saved adapter reloaded and scored all three types. This establishes training and reload compatibility on that hardware; task quality and each worker's capacity still require independent evaluation and qualification.
Existing jobs
JevK5 and Kev identities, references and artifact formats remain supported. Keep their references installed while their jobs finish. A worker may supply multiple --reference arguments; it claims only the profiles separately approved by the operator. Never rewrite a JevK5 job or adapter as H2O.
H2O requires fresh worker qualification and runtime hashes. Existing JevK5 capacity/quality reports do not transfer. Hosted migration, release-serving isolation and rollback are described in the platform rollout guide.