Model architecture

Dohnuts accepts a state and independent questions and returns temperature-scaled probabilities over supplied candidates. The training and serving templates are shared. A request reuses its causal prefix and computes question suffixes in parallel. Frozen image features are encoded on a cache miss and shared. More questions still require more suffix compute.

Each sequence contains the state, question, candidate descriptions, and a reserved marker after each candidate. A shared scalar head scores the marker’s hidden state. A per-question softmax gives the decision distribution. Nominal candidate order is shuffled during training; binary and ordinal order are fixed. Candidate-order sensitivity is measured on held-out data.

Fixed recipe

The Qwen3.5 adapter freezes the base and vision encoder and trains rank-8 language LoRA plus the scorer. The training recipe fixes precision, image and sequence budgets, optimizer, learning rates, sampling, checkpoint selection, and calibration. recipe.py is the executable specification. Run metadata records all values and data hashes, including values that callers cannot change.

RLCDConfig(sigma=0.3, ce_weight=1.0) is the complete RLCD configuration surface. sigma controls exploration in logit space; ce_weight controls the joint cross-entropy term. Four samples, the proper-scoring reward, ordinal distance penalty, baseline estimator, and numerical safeguards are fixed. --steps controls the total update budget.

The release path merges LoRA before calibration and final evaluation. Loading the compact checkpoint reconstructs that same merged model from the pinned base. The caller does not choose merge state, precision, attention backend, LoRA rank, or calibration mode.

Model adapters

Qwen35Adapter contains the backbone-specific code. Another backbone supplies an adapter object to DecisionModel, DecisionCollator, train, and final_evaluation; Predictor.from_checkpoint(..., adapter=adapter) validates its identity against the saved artifact.

Adapter responsibility	Contract
name, base_model, marker, image_prefix	Stable identity, one reserved candidate token, image prompt syntax
processor, load, hidden_size	Load the pinned processor/backbone and size the shared scorer
adapt_language, merge, freeze_vision	Apply the training and deployment lifecycle
batch_inputs, shared_image_inputs	Preserve all candidates and process shared images once
forward	Return candidate-addressable hidden states and the shared-prefix position offset

There is one supplied adapter. Supporting a different backbone requires its implementation and the same checkpoint/quality acceptance; an arbitrary model path alone does not establish compatibility. Model-specific tokenization and kernels belong here, without branching the RLCD objective or public decision API.

The inference reference describes question types, response fields, and input limits. Bub integration exposes the same decision API to agents.

Behavior and regression tests

tests/acceptance.py loads the exported checkpoint and exercises the real prediction and Bub interfaces. It checks text and image decisions, candidate probabilities, independent requests, reload, and input limits. The training workflow runs this acceptance on its exported weights.

The CPU tests in tests/test_*.py cover reported metric meanings, temperature calibration, the pinned RLCD objective and gradients, and the regression where identical screenshots crossed splits under different source IDs. Run them when changing the corresponding behavior:

make test

Tests assert observable behavior or a known failure case. Internal allocation, helper structure, log keys, file counts and exact error wording are not contracts. Quality evaluation separately measures held-out accuracy, calibration, candidate-order sensitivity and image dependence.