v0.3.3

Fixed

• Fresh installs now constrain datasets>=4.8.5,<5 so Toponymy does not pull an old
  datasets build that fails with current pyarrow releases.

v0.3.2

Changed

• Configuration loading is now strict for CLI, Python overrides, and run variants: unknown sections or keys fail early instead of being accepted silently.
• Toponymy outputs now use only canonical snake_case layer columns such as topic_layer_0_id and topic_layer_0_label; the working-layer topic_id and Name columns remain.
• The root ads_bib package now exports only the primary public entry points; lower-level stage helpers stay importable from their own modules.
• Hugging Face configuration now uses HF_TOKEN as the single public token environment variable.

Removed

• Removed the intermediate curation.cluster_targets config key. Use curation.layered_clusters_to_remove for layer-aware Toponymy curation.
• Removed legacy Topic_Layer_X output columns from new Toponymy runs.
• Removed public HF_API_KEY and HUGGINGFACE_API_KEY token names from docs, bootstrap, and doctor hints.

v0.3.1

Added

• ads_bib.run(from_run=...) now mirrors CLI run variants from Python, including dotted overrides and run-local artifact reuse.

Improved

• Toponymy topic-map layer colors now spread across the full palette for clearer hierarchy inspection.
• Curation docs now clarify that cluster IDs are run-local and that removals should be applied through a variant from the inspected run.

Fixed

• curation.clusters_to_remove is now validated when configs load; scalar values such as 7 fail early with a hint to use [7].
• CLI run and doctor now report config validation errors without a traceback.

v0.3.0

Added

• External source input support for starting runs from prepared publication and reference Parquet files instead of an ADS search/export step.
• Repository utility scripts for preparing source input from Semantic Scholar and INSPIRE exports.

Changed

• The local_gpu preset is now the recommended Colab/GPU road: TranslateGemma for translation, Qwen3 embeddings, and Qwen3 topic labels.
• pipeline.ipynb is now a preset-driven Colab quickstart that loads local_gpu, sets only the example query/run context, and uses notebook-friendly output.
• The previous Gemma embedding/labeling notebook has moved to notebooks/pipeline_gemma_experiment.ipynb as a non-primary experiment.

Improved

• Local Transformers translation now batches TranslateGemma calls, retries smaller batches on memory pressure, and applies dynamic generation limits inside the configured maximum.
• Local model stages now share cleanup helpers so GPU memory is released between translation, embeddings, and topic labeling.
• Notebook runs now pre-load the local models through package paths and keep normal progress output cleaner.

Fixed

• Local topic labeling now uses chat-template based interaction for local chat models, avoiding instruction text leaking into BERTopic labels.
• Optional TorchCodec import failures are handled for the local Hugging Face stack used in Colab.
• Topic-map rendering now falls back cleanly when datamapplot cannot form cluster boundary polygons.
• The colorspacious SyntaxWarning emitted during topic-map setup is filtered at the runtime logging boundary.

v0.2.0

Added

• CLI run variants via ads-bib run --from-run <run-id-or-path> --set ..., including automatic stage planning, --dry-run, downstream artifact hydration for visualization/citation-only variants, and optional variant provenance in run_summary.yaml.

Changed

• Final dataset bundle exports now clean publication/reference keys, prune dangling reference IDs, and remove placeholder or duplicate author UIDs before writing public Parquet outputs and the dataset manifest.
• New runs now use a modular artifact layout under runs/<run_id>/data/, with run-local stage restart points (search, export, translated, tokenized, and) plus final dataset and citations outputs; artifact_layout_version: 2 is recorded in run_summary.yaml.
• Translated and tokenized snapshots now carry metadata fingerprints so changed-config variants do not reuse stale source/config combinations, and enabled AND runs let ads-and validate its own cache metadata instead of loading disambiguated snapshots directly.
• OpenRouter embedding defaults now use larger documented batch sizes, the OpenRouter preset pins Toponymy-internal embeddings to Qwen3, and OpenRouter retries now fail fast on non-retryable request/auth/payment errors.
• Tokenized snapshot metadata now stores AND source fingerprints so validated AND cache hits can avoid recomputing expensive frame fingerprints on future runs.

v0.1.1

Fixed

• Toponymy fitting now avoids a large fixed-width Unicode array allocation that could cause memory errors on large corpora.
• Notebook and session resume now load translated, tokenized, and author-disambiguated snapshots even when earlier-stage frames are already in memory.

Changed

• The OpenRouter notebook example now uses Gemini Flash 3 for translation and topic labeling, with Qwen3 embeddings.

v0.1.0

Added

• Shared package runner in ads_bib.pipeline with structured PipelineConfig, named stages, and reusable stage functions.
• Thin CLI batch entrypoint: ads-bib run --config ... with optional --from, --to, --run-name, and --set overrides.
• Notebook adapter in ads_bib.notebook with NotebookSession and package-side config invalidation.
• Native huggingface_api translation path via huggingface_hub.AsyncInferenceClient.
• Official packaged runtime presets exposed via ads-bib run --preset ... and ads-bib preset write ....
• Workspace bootstrap and stage-aware doctor commands for first-run setup and preflight validation.
• Offline HF provider smoke coverage plus env-gated live HF smoke tests for translation, embeddings, and BERTopic labeling.

Changed

• Base ads-bib installs now own the official runtime stack; only non-default algorithm overrides remain behind the umap and hdbscan extras.
• Pin datamapplot to >=0.6.4,<0.7: 0.7.x changed the selection_handlers layout and breaks ads_bib.visualize until imports are updated.
• GitHub Actions now install only the active base contract plus test, umap, and hdbscan; removed references to historical extras and install profiles.
• pipeline.ipynb now uses explicit section dicts plus NotebookSession; it no longer owns config assembly, invalidation, globals() syncing, or START_STAGE / STOP_STAGE.
• Stage slicing remains a CLI/YAML concern; notebook reruns are driven by executing the corresponding stage cell.
• Notebook stage cells are now strict and no longer auto-chain earlier stages such as translate -> export.
• Fresh in-memory notebook state now takes precedence over same-stage translated/tokenized/disambiguated snapshots when a config change invalidates later stages.
• Run config snapshots are now serialized from structured pipeline config instead of raw notebook globals.
• Prompt selection now supports topic_model.llm_prompt_name with package-side resolution and .env fallbacks for ADS/OpenRouter secrets.
• Tokenization defaults now use en_core_web_md rather than en_core_web_lg.
• AND integration remains optional, but the active path is now the source-based external adapter rather than a placeholder notebook contract.
• run_pipeline() remains the dependency-aware batch path; notebook stage execution now has intentionally different UX semantics.
• Runtime output is now frontend-aware: CLI runs use compact stage-first console output, notebook runs stay slightly more explanatory, and raw third-party stdout/stderr is redirected into runs/<run_id>/logs/runtime.log.
• Nested progress-bar noise was reduced so normal runs show at most one primary progress bar per stage.
• huggingface_api embeddings now use the native Hugging Face async client instead of LiteLLM, while BERTopic labeling keeps BERTopic's LiteLLM adapter with normalized HF-native model ids.
• Pipeline config preparation now injects HF_TOKEN into translation, embedding, and BERTopic labeling configs when huggingface_api is selected.
• CLI runs now persist run_summary.yaml just like notebook runs, including partial/failure status metadata.
• OpenRouter and Hugging Face chat translation now share the same centralized scientific translation prompt contract.
• Official runtime roads now ship as four packaged generic presets accessed via CLI rather than repo-root YAML files.
• Stable local presets now pin only GGUF model families that are validated against the baseline ADS_env runtime; the CPU labeling preset uses Qwen/Qwen2.5-0.5B-Instruct-GGUF instead of unsupported qwen35 variants.
• Base runtime dependencies now include the provider and topic stack needed by the four official roads; huggingface-hub remains part of that default install.
• Hugging Face API key resolution now accepts HF_TOKEN, HF_API_KEY, and HUGGINGFACE_API_KEY.
• Core runtime dependencies now include pyarrow and networkx, and translation now validates the openai dependency for OpenRouter before execution.
• Packaging extras no longer expose the obsolete translate-local / translate-api names; the remaining extras are test, umap, and hdbscan.

Docs

• Site configuration lives at zensical.toml in the repository root; build and preview use zensical ... from the root (including GitHub Actions).
• Package_ToDo.md remains maintainer-local; repository cleanup no longer depends on a versioned backlog file in the public tree.
• Removed CLAUDE.md; repository engineering rules and conventions live in AGENTS.md only.
• Public docs and metadata now position the installed package and CLI as the primary runtime path, with pipeline.ipynb documented as an optional GitHub companion.
• AGENTS.md architecture notes now record the notebook-session adapter and the source-based AND step.
• README/runtime docs now document ads-bib run as the happy path, keep bootstrap and doctor as support commands, and treat huggingface_api as a full official road across both topic backends.