Explore ml data

Explore ML Data

Understand the dataset before designing a model. One project-level EDA per workspace: an executable data/eda.py, a persisted data/eda.md narrative, rich data/eda_<table>.html reports, and a short JOURNAL section that links them. The findings feed the baseline design note's learner / splitter / metric choices.

Next-step pointers — where you go after this skill

You came here for…	→ next
Bootstrap, before the first baseline	→ back to iterate-ml-experiment § 0; the EDA findings inform the auto-drafted 01_baseline.md
User free-text ("explore the data")	→ surface the findings; no further dispatch unless the user asks to model
Re-understand a changed data source	→ re-run, overwrite data/eda.*, refresh the JOURNAL EDA section

Always re-emit the Pre-flight checklist with evidence before declaring the turn done.

Where this sits in the loop

EDA is a bootstrap-time gate (G-EDA) owned by this skill and fired by iterate-ml-experiment § 0 before the baseline design note. Ordering matters: the dataset facts (class balance, datetime / group columns, missingness, cardinality) are exactly what justifies the splitter (G-CV-SPLITTER), the metric default, and the learner default. Running EDA after the model is designed defeats the purpose.

scaffold → JOURNAL → goal from data/README.md
   │
   └─► G-EDA (run | skip)  ◄── this skill
         │ run
         └─► data/eda.py → execute → data/eda.md + HTML + JOURNAL §EDA
   │
   └─► auto-draft 01_baseline.md  (cites the EDA findings)

Where things live — visual map

Two locations are kept separate: the raw data source (read-only, may live anywhere) and the EDA deliverables (always under <project>/data/).

Path	Durability	Who writes it	What it holds
raw data source (data/, raw/, an absolute path, external)	user-owned, READ-ONLY	the user	The dataset. EDA reads it; never modifies it. May be anywhere — not assumed to be data/
data/eda.py	Durable (committed)	This skill, once per workspace	The jupytext # %% EDA cells. Source of truth. Openable as a notebook for the rich view
data/eda.md	Durable (committed)	This skill (authored from the digest)	The prose narrative: findings + modelling implications that the baseline note cites
data/eda_<table>.html	Durable (committed)	data/eda.py via TableReport.write_html(...)	The rich, interactive skrub report per table — for the human
scratch/eda/eda.md	Ephemeral (gitignored), optional	run_cells.py when given a 2nd arg	Per-cell digest the agent reads. Same content as stdout
journal/JOURNAL.md § Data understanding (EDA)	Durable (committed)	This skill	2–4 line summary + link to data/eda.md

Mnemonic: the raw data is read-only and lives wherever the user keeps it; data/eda.py is source; data/eda.md + the HTML are the durable deliverables, always under data/; scratch/eda/ and stdout are the ephemeral run digest.

Read-only-against-raw-data contract

The central rule. Surfaced as the first Stop condition below.

Allowed — this skill writes ONLY (deliverables always under <project>/data/, created if absent):

• data/eda.py — the EDA script (created / overwritten in place).
• data/eda.md — the authored narrative.
• data/eda_<table>.html — the skrub TableReport pages.
• scratch/eda/ — the ephemeral digest.
• journal/JOURNAL.md § Data understanding (EDA).

Forbidden:

• Modifying, deleting, renaming, re-encoding, or "cleaning" the user's raw data files — wherever they live (data/, another folder, an absolute/external path). EDA reads them; it never rewrites them. Data cleaning is the pipeline's job (build-ml-pipeline), declared at fit time, not a one-off mutation.
• Writing anywhere outside the five paths above — no src/<pkg>/ edits, no reports/ writes, no new experiment files.
• Designing the model: no skore.evaluate(...), no project.put(...), no learner selection here. EDA informs those; it does not make them.

Stop conditions — read before anything else

• Read-only against the user's raw data. See § Read-only- against-raw-data contract. data/eda.py reads the raw files (wherever they live) and writes only the data/eda.* deliverables.
• Deliverables always under <project>/data/; the raw source is separate. Write data/eda.py / data/eda.md / data/eda_<table>.html under <project>/data/ (create the folder if absent). The raw data the script reads may live anywhere (data/, another in-repo folder, an absolute or external path) — decouple the two: a RAW = <LOAD_RAW_DATA> source vs an EDA_DIR output. Never assume the raw data is in data/.
• EDA precedes model design (G-EDA). In bootstrap, the gate fires before journal/01_baseline.md is drafted. It is binary: run (place + execute data/eda.py, write the deliverables) or skip (record Status: skipped — <date> in the JOURNAL section and proceed). Do not silently bypass — fire the AskUserQuestion. Free-text "go fast" / "quick baseline" does NOT resolve it.
• Agent feature required to execute. The cell runner needs ipython. If it is missing and the user chose run, STOP and delegate to python-env-manager § "Agent feature" (G-AGENT-FEATURE). Do NOT type pixi add ... ipython yourself; do NOT fabricate EDA output with hand-written print()s. If the user declines the agent feature, fall back to the skip path (record Status: skipped) — never loop between run and install.
• Symbol from memory is forbidden. Any skrub / pandas / polars symbol (TableReport, TableReport.json, write_html, column_associations, the tabular reader, …) must come from python-api this turn. Cache hits under scratch/api/<lib>/<version>/ count; inline memory does not. TableReport.json()'s key names are not formally documented and drift across skrub versions — confirm them via python-api and parse defensively (.get(...)).
• Library-agnostic — read facts off skrub, not pandas/polars. The workspace may use pandas OR polars (G-TABULAR), whose summary methods differ (select_dtypes doesn't even exist in polars). The structured facts come from skrub (TableReport(...).json(), column_associations), which accept both. The ONLY library- specific line is RAW = <LOAD_RAW_DATA>. Do not write df.isna()/df.nunique()/df.select_dtypes(...) etc.
• skrub.TableReport for dataframe overviews. Every table gets a TableReport(RAW, title=..., verbose=0) written to data/eda_<table>.html (the user-facing artifact) AND read via .json() for the digest. verbose=0 keeps progress prints out of the digest.
• Never end a cell on a bare TableReport. Outside a notebook, repr(TableReport(df)) is the useless <TableReport: use .open() to display>. Use report.write_html(...) (a statement) for the HTML, and end cells on text-friendly expressions (RAW.shape, a dict/list built from report.json(), skrub.column_associations(RAW)) so the digest carries real values. Mirrors audit's .frame() rule.
• Never gitignore the whole data/; ask about the inputs. The deliverables live in data/ and must stay committable, so the whole data/ folder must never be in .gitignore. If the raw inputs should be kept out of git (large / local-only), fire an AskUserQuestion offering to ignore specific input patterns (e.g. data/raw/, data/*.parquet) — default: don't. Then verify the deliverables are tracked (git check-ignore data/eda.md must return nothing). Never auto-edit .gitignore — that is organize-ml-workspace's to write; surface the patch and ask.
• One project-level EDA. A single data/eda.py covers the whole dataset; multi-table data gets one TableReport cell per table inside that one file (run the target/structure cells on the target-bearing table). No eda_v2.py, no per-experiment EDA files, not part of the four-way stem pairing. Re-understanding overwrites data/eda.py in place.
• Don't design the model here. No splitter pick, no metric pick, no learner pick. Record implications in data/eda.md; the picks happen in their owning gates (G-CV-SPLITTER, the baseline note).
• Harness "no clarifying questions" hints do NOT waive G-EDA or G-AGENT-FEATURE. Both fire regardless.
• Post-hoc audit — required before ending the turn. Walk every pre-flight row; surface unfilled Evidence cells explicitly.

Forbidden shortcuts

Shortcut	Why it's wrong
Design the baseline first, EDA "later if there's time"	Inverts G-EDA. The point is to justify the modelling choices before making them. EDA runs first in bootstrap
End a cell on a bare TableReport(df) to "show the report"	Outside a notebook that repr is <TableReport: use .open() to display> — zero signal in the digest. Use write_html(...) + a text summary built from report.json()
print(...) instead of a bare summary expression	The runner captures bare last-expressions via result.result; print(...) lands in stdout and is harder to scan. Use bare expressions
Use pandas/polars methods (df.isna(), df.nunique(), df.select_dtypes(...)) for the summaries	Breaks on the other library (polars has no select_dtypes). Read the facts off skrub (TableReport(...).json(), column_associations) — agnostic to pandas/polars
Clean / impute / drop columns in data/eda.py and re-save the raw file	EDA is read-only against raw data. Cleaning belongs in the pipeline (build-ml-pipeline), applied at fit time for train/test consistency
Assume the raw data is in data/	The raw source may live anywhere; only the deliverables are pinned to data/. Set RAW = <LOAD_RAW_DATA> to wherever the data actually is
Gitignore the whole data/ folder	The committed deliverables (data/eda.*) live there. Ignore only specific input patterns, and ask the user first
Run EDA without the agent feature by hand-writing the expected output	Fabricated EDA is worse than none. Missing runner → G-AGENT-FEATURE (install) or the skip path
pixi add ipython directly from this skill	Install is owned by python-env-manager. This skill requests via G-AGENT-FEATURE
Drop the authored data/eda.md and leave only the HTML	The .md carries the modelling implications the baseline note cites and the JOURNAL section links. Both are required
Invent column meanings not visible in the data	Report what the data shows. Domain semantics the user didn't state go in an explicit "open questions" list, not as asserted fact
Forget the JOURNAL § Data understanding update	The section is the index entry; without it later sessions can't find the EDA. It is part of "done"

Pre-flight — emit before any write or execution

Pre-flight (explore-ml-data):
- [ ] Trigger: bootstrap-G-EDA | user-request | data-changed
      Evidence: caller + rule that matched
- [ ] Detection: EDA already present? data/eda.md + JOURNAL §EDA
      Evidence: ls / Glob on data/eda.md + Read JOURNAL §EDA
                | "n/a — first EDA"
- [ ] G-EDA resolved: run | skip
      Evidence: AskUserQuestion id=<id>, answer=<run|skip>
                | user free-text quote turn N
      If skip: JOURNAL §EDA records "Status: skipped — <date>"; STOP here.
- [ ] Tabular library known (G-TABULAR): pandas | polars
      Evidence: JOURNAL.md Status (Workspace decisions) | AskUserQuestion
                via data-science-python-stack
- [ ] Raw data located (may be outside data/): <paths / loader>
      Evidence: ls / Glob on the data location + the RAW load call placed
                in data/eda.py | user-quoted path turn N
- [ ] data/ not gitignored as a whole; deliverables will be tracked
      Evidence: `git check-ignore data/eda.md` returns nothing
                | AskUserQuestion id=<id> on ignoring specific inputs
                | "n/a — no .gitignore yet"
- [ ] Agent feature available (run path only):
        `pixi run -e agent ipython -c "print(0)"` exit 0
      Evidence: tool output | JOURNAL.md Status `agent feature: installed`
                Missing → STOP, delegate to python-env-manager G-AGENT-FEATURE
                (decline → fall back to skip path)
- [ ] python-api consulted for symbols used:
        skrub.TableReport, TableReport.write_html, TableReport.json,
        skrub.column_associations, the tabular reader (load cell only)
      Evidence: Read/Write scratch/api/<lib>/<version>/<topic>.md (this turn)
                | "n/a — cache hit + Read this turn"
- [ ] Template copy + substitution decided:
        <pkg> → package name from src/<pkg>/
        <LOAD_RAW_DATA> → the real loader, pointing wherever the data lives
        <TARGET_COLUMN> → the target (from goal / data/README.md), or n/a
        <table> → short slug per table for eda_<table>.html
      Evidence: Read templates/eda.py this turn before Write data/eda.py
- [ ] Execution command shape confirmed:
        pixi run -e agent python \
          .agents/skills/audit-ml-pipeline/scripts/run_cells.py \
          data/eda.py [scratch/eda/eda.md]
      Evidence: command emitted before running
- [ ] Deliverables written: data/eda.md (prose + implications),
        data/eda_<table>.html (≥1), JOURNAL §Data understanding
      Evidence: Write of each | "n/a — skip path"
- [ ] Pre-flight re-emitted with evidence before final message.
      Evidence: this checklist appears in the end-of-turn summary.

EDA file contract — overview

data/eda.py is jupytext percent format (# %%), executed by the shared runner. Template: templates/eda.py. Full cell-by-cell anatomy with right / wrong shapes: → references/cell_anatomy.md.

Substitutions

Placeholder	Replaced with
<pkg>	The importable package name (from src/<pkg>/); used for from <pkg> import PROJECT_ROOT (only to locate EDA_DIR = PROJECT_ROOT / "data")
<LOAD_RAW_DATA>	The real load of the raw file(s), pointing wherever the data lives (in data/, another folder, an absolute path, or external). Uses the workspace tabular lib (pandas/polars); skrub accepts both. The one library-specific line
<TARGET_COLUMN>	The target column name (from the goal / data/README.md), or remove the target cell if unsupervised / unknown
<table>	A short slug per table for the HTML filename (eda_<table>.html) — for a single table use the dataset name

Cell sequence (what each cell does)

Brief outline; concrete examples → references/cell_anatomy.md.

1. Module docstring (markdown) — what this file is, the read-only-against-raw-data rule, raw-vs-deliverables split, how it is executed.
2. Imports + paths (code) — import json, import skrub, from <pkg> import PROJECT_ROOT, EDA_DIR = PROJECT_ROOT / "data" (+ EDA_DIR.mkdir(parents=True, exist_ok=True)). No pandas/polars import here.
3. Load raw data (code, bare expression) — RAW = <LOAD_RAW_DATA> pointing wherever the data lives; end on RAW.shape.
4. Per-table overview (code) — report = skrub.TableReport(RAW, title=..., verbose=0); report.write_html(EDA_DIR / "eda_<table>.html"); then summary = json.loads(report.json()) and end on a dict/list of per-column dtype / null / cardinality facts. One such cell per table.
5. Target analysis (code, bare expression) — pick the target's entry out of summary["columns"]; it carries value counts (classification) or a distribution summary (regression). Drives the metric default and whether the splitter should stratify.
6. Structure signals (code, bare expression) — datetime columns (from skrub's inferred dtypes, catches string dates) and high unique-ratio id/group columns. Drives the G-CV-SPLITTER choice (TimeSeriesSplit / GroupKFold).
7. Associations (code, bare expression) — skrub.column_associations(RAW) to flag strong predictors and possible leakage.
8. End (markdown) — reminder that the agent now authors data/eda.md + the JOURNAL section from this digest.

write_html(...) is load-bearing on the overview cells (the human artifact). verbose=0 and the bare report.json()-derived expressions are load-bearing for a clean, library-agnostic digest. For multi-table data, run cells 5–7 on the target-bearing table; for very large data, load a row sample (see references/cell_anatomy.md).

Execution contract — one command

pixi run -e agent python \
  .agents/skills/audit-ml-pipeline/scripts/run_cells.py \
  data/eda.py

The runner (shared with audit-ml-pipeline) streams the digest to stdout — the agent reads it directly from the bash tool output. Pass a second arg scratch/eda/eda.md to also write the digest to a file. For non-pixi workspaces, swap the activation prefix per python-env-manager § "Agent feature".

This skill ships no runner of its own — there is no explore-ml-data/scripts/. Always invoke the shared audit-ml-pipeline/scripts/run_cells.py at the path above; don't look for or fork a local copy.

Prerequisites for the run path: the workspace package must be importable (from <pkg> import PROJECT_ROOT — editable install done during scaffold) and skrub installed (Tier 1). If either import fails, the digest shows the ImportError; route to python-env-manager for the missing piece rather than working around it.

Re-execution semantics

• A changed / added data source → overwrite data/eda.py, re-run, re-author data/eda.md + HTML, refresh the JOURNAL section.
• scratch/eda/ is overwritten on every run. The durable record is data/eda.py + data/eda.md + git history.

Authoring data/eda.md

After the run, read the digest and write data/eda.md from templates/eda.md. It is prose, grounded in the digest — no invented facts. Required sections:

• Dataset at a glance — tables, rows × columns, target.
• Per-column findings — dtypes, missingness, cardinality highlights, anything surprising.
• Target — balance / skew; class counts or distribution summary.
• Structure — datetime ordering, groups / ids (or "none found").
• Associations — strong feature↔target / feature↔feature links; flag possible leakage explicitly.
• Modelling implications — the payoff section. Translate findings into candidate picks the baseline note will weigh: e.g. "imbalanced target → StratifiedKFold + look at ROC-AUC / PR-AUC, not accuracy"; "user_id repeats across rows → consider GroupKFold"; "timestamp present → TimeSeriesSplit if forecasting". These are implications, not decisions — the gates own the picks.
• Open questions — domain ambiguities for the user to confirm.

Link each data/eda_<table>.html from the relevant section.

JOURNAL § Data understanding (EDA)

iterate-ml-experiment's JOURNAL.md carries a top-level ## Data understanding (EDA) section (placed right after ## Status). This skill owns its content:

## Data understanding (EDA)

- **Status:** done — <YYYY-MM-DD>   <!-- or: skipped — <YYYY-MM-DD> -->
- **Summary:** <2–4 lines: dataset shape, target balance/skew, the
  one or two findings that most shape the modelling choices>
- **Report:** [data/eda.md](../data/eda.md)

Keep it to a few lines — it is an index entry, not the report. The detail lives in data/eda.md. On the skip path, only the Status: skipped line is required.

Dispatching in and out

Called from

Caller	When
iterate-ml-experiment § 0 bootstrap	Automatic; G-EDA fires before the baseline design note
User free-text	"explore the data", "do an EDA", "profile the dataset" — resolves directly

Calls into

Callee	Why
python-env-manager § Agent feature	When ipython is missing on the run path — G-AGENT-FEATURE
python-api	Every skrub / pandas / polars symbol. Cache hits first
data-science-python-stack	G-TABULAR (pandas / polars) if not yet recorded; skrub TableReport reference
python-code-style	After writing data/eda.py — ruff format / check + contextualize the comments to this dataset (strip any leftover workflow/process prose)

What this skill does NOT do

• Design, select, or evaluate a model (build-ml-pipeline / evaluate-ml-pipeline / iterate-ml-experiment).
• Pick the CV splitter or metric — it only surfaces the evidence for those picks.
• Edit src/<pkg>/ or the experiment / audit files.
• Clean, transform, or re-save the user's raw data.
• Install ipython / pyright (python-env-manager owns).
• Open or write the skore Project.
• Render commits or PRs.

Companion skills

Skill	Relationship
iterate-ml-experiment	Caller. § 0 fires G-EDA before the baseline note; the EDA findings seed the note's Method / Risks
audit-ml-pipeline	Owns the shared cell runner scripts/run_cells.py this skill executes; same bare-expression discipline
organize-ml-workspace	Workspace layout; data/ is user-owned — this skill is the one exception that writes data/eda.* into it
python-env-manager	Agent feature install (G-AGENT-FEATURE). This skill requests; that skill installs
python-api	skrub / pandas / polars symbol lookups. Cache hits first
data-science-python-stack	G-TABULAR; skrub TableReport is catalogued there
python-code-style	ruff after writing data/eda.py

Templates and assets

• templates/eda.py — the data/eda.py skeleton. Copy + substitute; don't rewrite from memory.
• templates/eda.md — the data/eda.md report skeleton.

The cell runner is not owned here — it is audit-ml-pipeline/scripts/run_cells.py (shared). Don't fork it.

References (load on demand)

• references/cell_anatomy.md — concrete cell examples (right / wrong shapes), the TableReport repr trap, the full cell sequence, and how each finding maps to a downstream gate.