LabelLens

Review noisy text labels from your terminal.

Your LLM, rules, or weak model labeled the dataset. Some fraction is wrong in ways you only catch by reading the rows. Options today:

• Spreadsheet — no record context, no prediction metadata, no real keyboard flow.
• Label Studio / Doccano / Prodigy — server, browser, accounts. Heavy for a 30-minute job over SSH, and overkill when you already have predictions and just need to review them.
• A one-off CLI you vibe-code per dataset — works once, then you rewrite it next time.

LabelLens is the missing keyboard reviewer in the middle. Accept, reject, relabel, ask an assistant, export.

Local-first, runs over SSH, ships as a single Bun-compiled binary. State lives next to the source JSONL; the dataset never leaves the box unless you turn on the LLM assistant.

Status: v0.3. See PRD.md and docs/adr/ for design.

Supported task types

Current support is configured with task. See Review task types for examples and data shapes. Planned task types are tracked in the roadmap.

Task type	Status	Use it for
classification	Supported	One label per record, such as intent, topic, or category review.
boundary	Supported	Document or line segmentation where surrounding context matters.
multi-label	Supported	Records that can carry multiple labels via toggle-style review.
extraction	Supported	Form-style correction of structured fields, without span editing.
Pairwise / preference	Planned later	LLM output comparison and evaluation workflows.
NER / span review	Planned later	Character-level span correction; deferred because terminal span editing is the hard part. See PRD §19.

Install

curl (macOS arm64/x64, Linux arm64/x64):

curl -fsSL https://raw.githubusercontent.com/somus/label-lens/main/install.sh | sh

Pin a version with LL_VERSION=v0.1.2, override paths with LL_PREFIX / LL_BIN_DIR. Each release ships SHA256SUMS.txt; the installer verifies before extracting.

npm fallback (containers, non-shell environments):

npm install -g label-lens

60-second quickstart

labellens init data.jsonl   # infer schema, write labellens.config.json
labellens                   # open the review screen

In the TUI (default simple preset — arrow keys; see docs/reference/keybindings.md for the full table covering both simple and vim presets):

• a accept · r relabel · 1–9 quick-relabel · x reject · s skip
• m mark record · :bulk-accept / :bulk-relabel <label> / :bulk-reject / :bulk-skip act on all marked
• ↓ / ↑ navigate · → / ← cycle queues
• i LLM assistant (configures on first press) · t stats · ? help · q quit

When done:

labellens export jsonl       # write reviewed dataset
labellens export stats       # Markdown summary

Docs

Tutorial	Full 5-minute walkthrough.
How-to guides	Configure the assistant, work with queues, bulk-relabel, export, migrate labels, run over SSH, use Ollama locally.
Reference	Config schema, keybindings, queue grammar, CLI flags, output formats.
Explanation	Domain model, why skipped is its own state, audit semantics.
Roadmap	What's deferred past v0.3.
labellens guide	Print the tutorial offline (SSH-friendly).
labellens --help	Quick reference printed to stdout.
man labellens	Man page (installed by curl-installer).

Project layout

• PRD.md — product spec.
• CONTEXT.md — domain glossary.
• docs/ — user-facing documentation (index).
• docs/adr/ — architecture decision records.
• AGENTS.md (alias CLAUDE.md) — orientation for AI agents working on this repo.

License

MIT