v0.2.4

0.2.3 - 2026-09-11

Release Notes

Fixed bugs in v0.2.3 of the ITI implementation.

v0.2.2

0.2.2 - 2026-09-10

Release notes

• Artifact-first steering recipes, deterministic selection, and checksummed artifact bundles improve reproducibility and auditability.
• Versioned adapter capability declarations make supported representation surfaces explicit and fail closed for unsupported controls.
• LAT now follows RepE-style pair handling, and Hugging Face Qwen3/Qwen3-MoE models are supported.
• Hardened artifact I/O, expanded CI, and broader regression coverage improve reliability.

Added

• rs.recipes.directional_ablation(...) and rs.recipes.gated_direction(...) for composing common artifact-backed steering plans.
• repsteer.selection with deterministic holdout splits, JSON-only candidates, max/min selection, grid search, and checksummed replayable SelectionReport records.
• ArtifactBundle with role-keyed components, nested artifact verification, provenance, optional selection evidence, and component-level compatibility diagnostics.
• Versioned AdapterCapabilities, SampleMappingCapability, and ModalityMappingCapability declarations for text and VLM adapters.
• Qwen3Adapter support for standard Qwen3 decoder-only and Qwen3-MoE architectures.
• CPU CI across Python 3.10–3.14, pre-commit checks, and expanded tests for artifacts, selection, bundles, adapters, hooks, recipes, and multimodal runtime behavior.

Changed

• LAT now supports deterministic pair shuffling, explicit pair_signs, configurable shuffle_pair_order, pairwise sign voting, and RepE-compatible raw-difference scaling.
• Built-in adapters expose read-only capability records for tested residual, sample-mapping, and modality-mapping surfaces.
• Sequence-gate decisions are scoped to a single generation call, require one decision per batch item, and fail closed for targets without an explicit sample-mapping contract.
• Hugging Face Transformers, PyTorch, and safetensors are now base dependencies; SAE, VLM, and development dependencies remain optional extras.
• Artifact and selection records use stricter JSON schemas, checksum validation, and deterministic serialization.

Fixed

• Exact artifact compatibility now rejects explicitly known architecture mismatches.
• Artifact reads and writes reject symlinks, unsafe paths, unexpected files, malformed manifests, and checksum or digest mismatches.
• Sequence-gate caches are cleaned up on both successful and exceptional generation exits.
• Unsupported head-result and attention-bias surfaces now produce explicit fail-closed diagnostics instead of being approximated.

v0.2.1

0.2.1 - 2026-07-24

Release notes

• Native Hugging Face chat-template support for text instruction models and static-image VLMs.
• Structured messages, conversations, and conversation batches now work directly with generation, capture, learners, and sweeps.
• Template-specific inputs, rendering choices, and template hashes are recorded in capture fingerprints and artifact provenance.
• Safer instruction prompting through assistant-prefix defaults, duplicate special-token protection, and fail-closed validation for unsupported inputs.

Added

• Hugging Face chat-template generation for structured message mappings,
  conversations, and conversation batches. model.generate(messages=...) is
  supported alongside positional and prompt= inputs.
• chat_template_kwargs on generation, CaptureRequest, and contrastive
  learners for template-specific inputs such as tools, documents, or an
  explicitly selected template.
• Processor-preferred chat rendering for image generation, with tokenizer
  fallback, so instruction-style VLM prompts can use their native template.

Changed

• Chat generation adds an assistant generation prompt by default; capture and
  learner capture retain their explicit default of False.
• Rendered chat text is tokenized with add_special_tokens=False by default
  to avoid duplicating template-owned BOS/EOS/control tokens. An explicit
  tokenizer or processor option may still override that default.
• Capture cache keys include template kwargs, while learner artifact provenance
  records the resolved chat-rendering choice. Exact artifact compatibility now
  checks learned tokenizer-template hashes; sweep evaluation recognizes a
  one-message chat mapping as a prompt rather than generation keyword args.

Fixed

• Single message mappings and raw strings explicitly requesting a template are
  normalized before calling apply_chat_template; mixed raw/chat batches and
  disabled templates for structured messages fail with actionable errors.

v0.2.0

0.2.0 - 2026-07-23

Release notes

• SAE steering with portable feature artifacts, decoder-direction additions, and latent clamp/ablate operators.
• Sequence-level conditional gates and plan-composition diagnostics for controlled multi-intervention execution.
• Qwen2.5-VL and InternVL adapters with semantic vision/projector sites, image-token and patch selectors, and processor-driven generation.
• Multimodal provenance, compatibility checks, evaluation schemas, and SAE/VLM guides for reproducible steering workflows.

Added

• A backend-independent SAE protocol, lazy SAELens provider, custom/functional
  adapters, common SAELens hook-name-to-site inference, and portable
  SAEFeatureArtifact bundles.
• Supervised SAE feature ranking and causal-effect ranking, including a
  high-level contrastive capture path and supervised shortlisting before a
  user-supplied causal evaluator.
• SAE decoder-direction steering through the ordinary Add operator,
  direct-latent Clamp/Ablate, and residual-preserving
  SAEClamp/SAEAblate.
• ProbeGate, SAEActivationGate, cosine/callable gates, and logical gate
  composition. Explicit language-site gates are evaluated in prefill and
  cached per sequence for decode.
• Plan composition diagnostics for stable execution order, direction norms,
  same-target pairwise cosine, effective rank, condition number,
  orthogonalized bases, additive-fusion candidates, and non-commutative pairs.
• Qwen2.5-VL and InternVL architecture adapters with language, vision-residual,
  projector-input, and projector-output semantic sites.
• An adapter-produced ModalityMap plus serializable ImageTokens,
  ImagePatches, and normalized-XYXY ObjectPatches selectors. Qwen mapping
  accounts for dynamic static-image grids, spatial merge, and window order;
  InternVL mapping accounts for CLS offsets, pixel shuffle, and supported tile
  grouping.
• Static-image processor-driven generation, including native placeholder
  insertion for the unambiguous one-prompt/one-image case and fail-closed
  cardinality checks for batched or multi-image input.
• Processor and modality provenance in artifacts, exact preprocessing
  fingerprint compatibility, and a finite-only multimodal evaluation schema
  split into representation, causal-behavior, and capability/quality metrics.
• SAE/VLM guides and a compensatory-steering example that combines a
  vision-side prefill control and language-side decode control in one
  SteeringPlan.

Changed

• Hugging Face loading now selects an image-to-text model and AutoProcessor
  for recognized VLM configs; generation preserves its validated modality map
  across language decode.
• The compiler resolves and validates gate read dependencies, rejects
  non-language explicit sequence-gate sites, preserves priority/declaration
  order, and exposes composition data through diagnostics() and
  explain().
• A statically zero Constant or NormRelative schedule is compiled as a
  strict no-op: no target/gate hooks run and random-number-generator state is
  unchanged.

Fixed

• Sequence-gate caches are reset at every new generation and cannot leak into a
  first decode call backed by pre-populated KV state.
• Compiled plans cannot be rebound after their source model wrapper has been
  collected, even if Python later reuses the same object identity.
• Composition JSON no longer emits non-finite numbers or compares directions
  from different concrete targets.
• Exact artifact/processor compatibility includes preprocessing
  configuration, not only a declared processor id and revision.
• Public dtype= loading is translated to the Transformers 4.x-compatible
  torch_dtype keyword across the declared dependency range.
• InternVL ObjectPatches fails closed for multi-tile source images when the
  processor does not expose tile-to-original-image crop transforms;
  ImagePatches remains explicitly tile-local.

v0.1.0

• Initial static text activation-steering release.
• Hugging Face causal-LM runtime and semantic architecture adapters.
• ActAdd, DiffMean/CAA, PCA, LAT, and linear-probe learners.
• Typed sites, token selectors, schedules, operators, artifacts, and sweeps.

v0.0.0

Create python-publish.yml