references/sources.md
Dated official sources and verification
Review date: 2026-10-01. The July 2026 review remains historical evidence; package, source, dataset revisions, provider contracts and the local tools were rechecked for this refresh. Source inspection and mock transport tests are not authenticated provider or scientific validation.
Release and implementation
- PyPI package and official metadata: latest release remains 0.3.5, released 2025-07-16, Python >=3.10, beta classifier, MIT, dependencies, artifact hashes and trusted-publisher provenance. The 96,169-byte wheel was downloaded and its SHA-256 matched the metadata.
- Official repository,
latest release,
and default branch:
v0.3.5 and master
bd37a3129a2f98ee586f545a57b10b59496eedadremain unchanged. - Pinned release source and pyproject: source/API review targets this exact version rather than an unpinned branch.
- Generation CLI, inference CLI, and logger: parser flags and positional arguments, early help exit, swapped logger and hosted-wrapper constructor arguments, small-bank assertion and averaging bug.
- Task loader and prompt implementation: config fields, column JSON, sampling, return order, OOD aliasing, and template substitution. The shipped template rendered locally with synthetic examples.
- Model wrappers: SDK calls, cost table, local registration, model loading, retries, message mutation, block assumptions and error behavior. Executed hosted paths used mock HTTP responses and stubbed heavy import-only dependencies.
- Output serializer and SummaryInformation: JSON hypothesis-bank structure and stored statistics.
- Redis cache: cache keys, prompt/response storage and pickle serialization. No Redis server was started or contacted.
- PyTorch 2.4.1 release metadata:
no CPython 3.13 wheels; the upstream package's
torch~=2.4.0range still prevents the default repository interpreter from installing its full stack. - PyYAML release metadata and parser documentation: reviewed local parser pin advanced to 6.0.3. Valid YAML and rejection of duplicate keys, aliases, anchors, tags, dates and malformed input were exercised locally.
Datasets
- HypoBench GitHub repository
and default commit:
remain at
7e4bbc341ee90b7efaa607f67a81543cd68cdf2e. - Pinned dataset tree: the deceptive-review config and train/validation/test JSON hashes were rechecked; 800/300/500 rows and three cross-split duplicate groups reproduced. The optional OOD file was acquired only to verify the task's path contract.
- Official Hugging Face publication
and Hub metadata:
revision remains
613860dcbcda9e522a6163ee9edf78c261ebe4bb.
Provider contracts and privacy
- OpenAI Chat Completions:
request and response fields, legacy
max_tokens, choice/usage semantics and incompatibility with o-series models. The pinned OpenAI 1.40.3 SDK was exercised through HTTPX MockTransport; it sent the documented route and Bearer header. - OpenAI data controls and deprecations: endpoint/model-specific storage and availability must be checked at execution time. This review did not query an authenticated model catalog or account.
- Anthropic Messages:
top-level system field, content-block response, version header, token parameter,
and current temperature restriction. Anthropic 0.32.0 sent the expected route,
x-api-keyand API-version header in mock transport tests. - Anthropic model deprecations: retired model IDs and rejection of non-default temperature on newer models.
- Anthropic API retention and commercial retention: standard deletion policy, exceptions, ZDR eligibility, and covered-model 30-day requirements. No account-specific arrangement was verified.
- OpenAI 1.40.3 client source and Anthropic 0.32.0 client source: automatic credential lookup and base-URL environment overrides. These were also inspected in the installed pinned SDKs; no credential values were read.
Local wrappers and scientific scope
- Transformers 4.45.2 pipelines and installation/offline mode: model/path loading and offline configuration compatible with the pinned 4.45.x line. No model weights or local inference were tested.
- vLLM 0.6.2 LLM API: offline model, sampling and generation interfaces; reviewed alongside the package wrapper, not validated on GPU hardware.
- Hypothesis Generation with Large Language Models:
Zhou et al., NLP4Science 2024, DOI
10.18653/v1/2024.nlp4science-1.10; data-driven candidate generation and classification evaluations. - Literature Meets Data: Liu et al., version 3 dated 2025-01-08; HypoRefine and literature/data integration.
- HypoBench: Liu et al., version 2 dated 2026-02-10; seven real-world tasks, five synthetic families and 194 datasets. Publication metadata does not establish scientific validity of an individual generated hypothesis.
- HypoBench OpenReview record: current retrieval reached a browser challenge. No current review/acceptance status is asserted from that response.