dtflow

English | 中文

Inspect, clean, transform and ship LLM training data from one CLI; dt view is the interactive entry point.

A training file's row is not a row: it is a conversation, a preference pair, an instruction. Most tools see opaque JSON; dtflow sees the sample, and everything follows from that — a terminal browser that renders samples by their detected format, ~30 pipeable commands that filter on conversation structure, conversion and framework export with lineage, and an agent that can drive all of it.

_{dt view data.jsonl — one screen, one complete sample. Press P and what you just did becomes a dt filter … \| dt sort … pipeline.}

Try it in 30 seconds

pip install dtflow                      # or, without installing: uvx --from dtflow dt view data.jsonl

dt view data.jsonl                                              # browse (press ? for keys)
dt stats data.jsonl --schema                                    # nested schema: types, null rates, values
dt filter data.jsonl "turns(x) >= 2 and search(x, 'refund')" \
  | dt select - "id,turns=turns(x),last=x.messages[-1].content" \
  | dt sort - --by x.turns --desc | dt head - 5
dt transform data.jsonl --preset=openai_chat -o train.jsonl     # sharegpt / alpaca / dpo → OpenAI messages

Reads JSONL/NDJSON (also .gz), JSON, CSV/TSV, Parquet, Arrow and Excel. Every data command accepts - for stdin and writes to stdout when no -o is given, so everything composes. The interface is in English; dt lang zh switches it to Chinese. No data at hand? examples/ has small synthetic chat / ShareGPT / Alpaca / DPO files to try these on.

Why dtflow

• It knows what a training sample is. Generic table tools show messages as {3} or a truncated string. dtflow detects openai_chat / sharegpt / dpo / alpaca and works on the sample everywhere: the viewer renders one complete sample per screen (turns colored by role, tool calls formatted with bad JSON flagged, VLM images one key away), the row helpers turns(x) / roles(x) / calls(x) / search(x, 'refund') filter on conversation structure, presets convert between layouts, token-stats splits tokens by role.
• Looking and processing are one loop, in one language. Conditions everywhere are Python with the current row as x — no DSL, and the viewer compiles the same string as dt filter. The loop closes: P in the viewer turns what you did into dt filter … | dt sort …; | runs a shell pipe over the whole file with its output browsable right there; w exports the filtered subset with lineage. What you did while browsing becomes a script; a script's output comes back to the viewer with dt … | dt view -.
• An agent can drive all of it. stdout carries only data, exit codes are contractual, errors are structured JSON when stderr is not a terminal, --dry-run previews side effects, dt schema prints the command tree as JSON. dt install-skill installs the full reference into Claude Code or Codex — one command, and "clean this dataset" becomes a task your agent can actually do.

Underneath: filter / select / map / clean / dedupe / transform stream and never load the file; a bare field name or an unknown helper is a compile-time error, not an empty result with exit code 0.

The workbench in practice

Decontaminate — drop training rows that overlap your test set. The join key is a Python expression, so it matches on the first user turn rather than the whole JSON line; rows that differ only in metadata still get caught.

dt join train.jsonl test.jsonl --on "first_user(x)" --anti -o clean.jsonl
dt view clean.jsonl                    # eyeball what survived before training on it

Browse until you understand the data, then turn what you did into a pipeline.

dt view data.jsonl        # / search every value · F tick values to keep · f filter by expression
                          # P copies what you did as:  dt filter … | dt sort …
dt filter data.jsonl "turns(x) >= 4 and x.score > 0.7" | dt view -    # process, then look again

Convert and ship to a training framework.

dt transform shards/ --preset=openai_chat -o sft.jsonl   # a directory works too; sharegpt/alpaca/dpo → messages
dt validate sft.jsonl --preset=openai_chat               # schema check; --filter keeps valid rows only
dt token-stats sft.jsonl --model=gpt-4                   # token distribution per role
dt export sft.jsonl -f llama-factory                     # data + training config template (ms-swift, Axolotl too)

Or hand the whole job to your agent.

dt install-skill        # teaches Claude Code (or Codex) the full reference

Then: "drop rows whose last turn isn't the assistant, dedupe by first user message, export for LLaMA-Factory." The agent reads dt schema, previews with --dry-run, and runs the same commands above.

dt view: the eye of the loop

dt view <file> opens a master-detail browser: the table summarizes each sample (derived columns turns/roles/first_user/chars/calls plus your metadata), the detail pane renders the current row by format. JSONL, CSV and Parquet open; a 910k-row file opens instantly at ~90 MB RAM.

Main screen: pick a row, detail renders by role with code highlighting	/ full-file search: every value of every record (assistant replies included), hits in yellow
F value picker: Excel AutoFilter style, unique values + counts, keep what you tick	dpo comparison: chosen / rejected in different colors, format auto-detected

dt view data.jsonl                          # open, press ? for keys
dt view app.jsonl -100 -f                   # follow a growing JSONL (log mode)
dt view data.jsonl -w "turns(x)>=6" -s error   # start filtered + searched
dt sample data.jsonl 500 | dt view -        # pipe: inspect sampled / processed output

Keys	What they do
/ f F	Full-file search · expression filter (the same Python as dt filter) · Excel-style column value picker. They stack; r clears all
Enter n/N *	Zoom into the sample · jump field by field · jump only between search hits
w C P	Export the filtered subset (lineage written) · copy a reproducible dt view command · copy the same conditions as dt filter … | dt sort …
|	Run a shell pipe (dt filter - … | dt sort - …, jq works too) over the whole file and browse its output right there; r returns to the file
mouse	Click a header to open its value picker, drag column widths and the split, drag-select text to copy (over SSH and inside tmux too)

Against other terminal viewers: VisiData, tabiew, jless, fx and csvlens are generic table/JSON tools — none of them knows what a training sample is, none has mouse filtering, and opening a 155 MB / 910k-line JSONL costs them 550 MB–2.3 GB of RAM against dt view's 91 MB. Where they win (pivots, SQL, deep JSON folding) and the full comparison table: docs/view.md.

The commands

Every data command follows one contract: FILE may be - (stdin), a directory or a quoted glob; without -o data goes to stdout while progress goes to stderr; output format follows the extension. Conditions, keys and derived fields are Python expressions with the current row as x.

# look before you touch
dt stats   data.jsonl --schema                        # nested schema: types, null rates, sample values
dt describe data.jsonl "turns(x)" "x.score"           # quantiles + histogram per expression
dt token-stats data.jsonl --model=gpt-4               # token distribution per role
dt diff    a.jsonl b.jsonl --key=id                   # what changed between two versions

# shape
dt select  data.jsonl "id,n=len(x.messages)"          # project / rename / derive
dt map     data.jsonl "x.text = x.text.strip(); del x.debug"   # edit in place
dt explode data.jsonl --field messages --index-as turn         # one row per list element
dt group   data.jsonl --by "roles(x)" --top 10        # {"key","count","pct"}
dt sort    data.jsonl --by "len(x.messages)" --desc
dt sample  data.jsonl 1000 --by=meta.source           # stratified sampling
dt split   data.jsonl --ratio=0.9 --seed=42           # train/test files

# combine & clean
dt dedupe  data.jsonl --key=messages[0].content -i    # exact dedupe, in place
dt clean   data.jsonl --drop-empty=text --min-len=messages.#:2
dt concat  a.jsonl b.jsonl -o merged.parquet          # one file in = format conversion
dt run     pipeline.yaml                              # reproducible pipeline, steps = CLI commands

Rows whose expression fails (missing field, None > 0.5) don't abort the run — filter drops them, select sets the item to null, map keeps the row, and stderr prints one summary at the end; --strict turns the first failure into exit code 1. In a terminal, results preview as a 50-row table with the same columns as dt view; piped, they emit NDJSON.

Command reference: docs/cli.md · expressions and field paths: docs/expressions.md · pipelines: docs/pipeline.md.

Python API

The same operations as a chainable class; x.field attribute access works on nested dicts and lists.

from dtflow import DataTransformer, load_stream

(DataTransformer.load("data.jsonl")
    .filter(lambda x: x.score > 0.8)
    .transform(preset="openai_chat", user_field="q", assistant_field="a")
    .dedupe("messages[0].content")
    .save("train.jsonl"))

(load_stream("huge.jsonl")                 # constant memory for files that don't fit
    .filter(lambda x: x["score"] > 0.5)
    .save("filtered.parquet"))

Presets (openai_chat, alpaca, sharegpt, dpo_pair), schema validation, token counting with tiktoken or HuggingFace tokenizers, converters for LLaMA-Factory / ms-swift / Axolotl / HuggingFace datasets / OpenAI Batch, lineage tracking: docs/python-api.md.

Agent skill

dt install-skill                         # Claude Code (default)
dt install-skill --target codex          # Codex

After installing, /dtflow in Claude Code or $dtflow in Codex gives the agent the full reference.

Documentation

• Quickstart
• CLI reference
• dt view
• Expressions and field paths (includes the 0.9 migration table)
• Pipelines
• Python API
• Changelog

License

MIT