Migrating from the initial release

This page lists what changed since the first public version of the package (the state of the main branch before the repository was restructured). Behaviour that was a bug is described as fixed; breaking changes are marked.

Installation

• The core package no longer installs PyTorch, Transformers or the provider SDKs. Install the extras you need: huggingface, openai, ollama, google, deepseek, or models for all. A missing extra raises an ImportError that names it.
• scipy, nltk, protobuf, langchain, langchain-community and ipywidgets are no longer dependencies (ipywidgets is the notebook extra).
• Python 3.10 – 3.14 are supported.

Results

• Seeds are applied for local Hugging Face models, Ollama and Hugging Face endpoints (see Reproducibility). Results of earlier versions cannot be reproduced with seeds; re-run if you need reproducibility.
• The generation parameters of the default model (used when a configuration has no models key) are now applied; they were silently dropped, so answers contained the whole prompt.
• revision, cache_dir and huggingfacehub_api_token of local models are now honoured.
• The default prompt keeps its exact original text (a formatter had changed two spaces).

API changes

Before	Now
run() returned None	returns a RunSummary; new arguments max_concurrency, on_error, show_progress
failed calls were printed and skipped	logged, counted in the summary, one RuntimeWarning at the end (on_error="raise" stops)
an experiment could only be run once	experiments can be run repeatedly
experiment_from_file returned None for invalid files; experiment_from_dict raised an empty RuntimeError	both raise ValueError (FileNotFoundError for a missing file) with a useful message
experiments_from_files found nothing → RuntimeError	ValueError; files are loaded in alphabetical order
experiment.model_dump() (the documented trimmed dump never ran)	experiment.to_config() / export_to_file(); API keys are masked
api_key fields were plain strings	SecretStr; the DeepSeek and Google keys of a config are used; environment fallbacks work
seeds had to be strings	integers or numeric strings; non-integers are rejected when the config is loaded
add_model / add_persona with an existing identifier kept the old entry (model: half replaced)	replace it and warn
PromptRemovalCleaner.parse returned a dict	returns the cleaned string
PostprocessingPipeline printed progress, ignored default answer options	logs, falls back to the questionnaire's default options, adds a valid column, returns the frame
JSONLCallback, PrintCallback, PrintTableCallback dropped every answer (missing time argument)	work; JSONLCallback rows no longer embed all answers collected so far
import rupsycho took seconds	takes milliseconds; names and sub-packages load on first use
rupsycho.parser (copy of rupsycho.parsers.parser)	BasicParser is an alias of BasicCleaner
example_experiment_bfi	deprecated, use load_example_experiment("bfi", models={})

Removed: rupsycho.chains, rupsycho.utils.plots, rupsycho.models.instruction, rupsycho.models.profile, get_default_model, the unused alternative prompt templates in rupsycho.prompts, and the dead ExperimentExportMixin.model_dump.

New

rupsycho command, rupsycho.scoring, rupsycho.seeding, bundled examples (rup.list_examples()), ExperimentDocument.assemble_prompt, to_config, RunSummary, a configurator that keeps reverse-keying flags and answer weights on import, benchmarks and a much larger test-suite.