saturn-dissect 0.2.0

Dataset dissector: polars-native profiling, HTML + JSON reports, optional multi-provider LLM insight pass, live Flask viewer

saturn

Dataset dissector. Point it at a HuggingFace repo, a local file, or a slice of either and it produces a terminal summary, a self-contained HTML report, and a machine-readable JSON findings file. Stats pass is always free and deterministic. Language-model insight and topic clustering are opt-in.

Generic across domains: alt-text, Bluesky firehose, census tables, VQA annotations all work out of the box.

Primary use case: lukeslp/bluesky-alt-text, 404,841 image descriptions, profiled in 46 s, compared across the curated/firehose split in 18 s.

Install

python3.10 -m venv venv
source venv/bin/activate
pip install -e '.[nlp]'
# optional, enables true full-corpus language detection:
mkdir -p .cache/saturn
curl -sL -o .cache/saturn/lid.176.bin https://dl.fbaipublicfiles.com/fasttext/supervised-models/lid.176.bin

Python 3.10+ required.

Use

# Full-corpus profile of a HuggingFace dataset
saturn huggingface lukeslp/bluesky-alt-text

# Local file (CSV, JSONL, Parquet, SQLite)
saturn analyze path/to/data.csv

# Compare two slices of one dataset by a column value
saturn compare lukeslp/bluesky-alt-text --by source_mode \
    --label-a curated --label-b firehose

# Compare two independent sources
saturn compare hf://user/a hf://user/b

# Opt-in streaming sample when full load is too heavy
saturn analyze big-dataset.parquet --sample 5000

# Opt-in LLM insight pass (primary, with optional catfish critic)
saturn analyze data.csv --llm anthropic
saturn analyze data.csv --llm anthropic:claude-sonnet-4-6 --llm openai:gpt-4o-mini

Output

1. Terminal: row count, per-column type, null %, unique count, alerts (duplicates, high_skew, outliers, multilingual, near_unique, boilerplate, allcaps, one_word, url_heavy, and so on).
2. HTML report: one self-contained file, TOC with per-column charts and stats tables. Compare mode adds a "most divergent columns" summary driven by a composite score.
3. JSON findings: every number and string in the HTML, ready to feed a notebook generator or the live viewer.

Viewer

pip install -e '.[web]'
saturn serve --dir path/to/findings/ --port 5043
open http://127.0.0.1:5043

A WCAG 2.2 AA compliant live alternative to the static HTML report. Drop findings JSON files into a directory; refreshing the index picks up new runs without restarting. /api/findings/<id> returns the raw JSON for scripting.

Design

• Polars-native, full-corpus by default. 404K rows of 21-column Bluesky data profiled in under a minute. Opt-in --sample N for quick peeks on anything bigger.
• Bounded memory. duplicate_counter and vocab expansion get skipped on JSON-blob-shaped columns; near-unique columns skip value-counts that would return 400K singletons; vocab tokenisation is capped to a 20K-row subsample truncated to 500 chars per row.
• fasttext lid.176 for true full-corpus language detection when the model is present (~1M docs/sec). Falls back to a bounded langdetect sample otherwise.
• Two passes. A free deterministic stats pass (always runs) and an opt-in language-model insight pass (Phase 2, shipped).
• Schema inference uses absolute and relative cardinality: a 489-value column in a 404K-row corpus is categorical, not text, even though 489 > the absolute threshold.
• Compare mode is the dataset's feature. Diff two slices column-by-column; every delta (null drift, mean/length delta, entropy delta, top-value jaccard, language-mix jaccard) is both visible in the HTML and machine-readable in the JSON.

Real output

Running saturn compare lukeslp/bluesky-alt-text --by source_mode on the full 404,841-row corpus (12 to 18 s):

signal	curated (279K)	firehose (125K)	Δ
alt_text mean length	202 chars	281 chars	+79 chars
alt_text duplicate rate	8.5%	20.0%	+11.5pp
language jaccard	.	.	0.35 (wide divergence)
author_handle null	0%	100%	+100% (firehose is anonymised)
cursor duplicate rate	75%	17%	−58pp

The +79 chars on firehose vs curated was the non-obvious finding: the curated 489-account population writes shorter alt text than the broader stream. Worth a Concadia-style readability follow-up.

LLM insight pass (opt-in)

Pass --llm provider[:model] on analyze or huggingface to layer a narrated insight pass on top of the deterministic stats. Pass the flag twice and the second provider plays catfish critic: it reviews the first model's narrative against the same evidence and returns agree/disagree/partial. Insights land in both the HTML report and the JSON findings (key: insights). The pass fails open: provider errors or missing API keys never block the deterministic output; they are recorded in insights.errors and saturn exits 0.

Requires ~/shared/llm_providers on PYTHONPATH (the unified provider gateway). Supported providers: anthropic, openai, groq, gemini, mistral, cohere, xai, perplexity, huggingface, ollama.

Status

Phase 1 (stats + HTML + JSON), Phase 2 (--llm insight pass with catfish critic, analyze/huggingface), Phase 4 (compare mode, curated-vs-firehose diff), and Phase 5 (Flask viewer on port 5043) are all shipping. Phase 2.5 (compare-mode insights) and Phase 3 (BERTopic clustering) on the roadmap.

License

MIT. © Luke Steuber.