stamper

Human annotation tool for AI conversation datasets. Label conversations as pass / review / fail with a single keypress — no JSON editing, no setup.

Reads and writes JSONL, making it compatible with any evaluation pipeline, pandas, or database.

---

How it works

1. Load — reads a JSONL file of conversations (one per line)
2. Resume — if an output file already exists, conversations already annotated are silently skipped
3. Annotate — renders each conversation turn-by-turn in the terminal and waits for a single keypress
4. Write — appends the annotation immediately to the output file (flush on every keypress, safe on Ctrl+C)
5. Summarize (optional) — press s to trigger an LLM call that summarizes the conversation inline before you decide

Each annotation is written as the original JSON line with a human_annotation field added. Extra fields on conversations and turns are preserved as-is (passthrough).

---

Build

From source:

go build -o .bin/stamper ./cmd/

macOS note: downloaded binaries may be blocked by Gatekeeper ("not from a trusted developer"). Run once to remove the quarantine flag:

xattr -d com.apple.quarantine .bin/stamper

---

Run

# Output defaults to sampled_annotated.jsonl
stamper -i sampled.jsonl

# Explicit output file
stamper -i sampled.jsonl -o my_annotations.jsonl

# Without building
go run ./cmd/ -i sampled.jsonl

Flags

Flag	Default	Description
-i / --input	required	JSONL file of conversations to annotate
-o / --output	{input}_annotated.jsonl	Annotation output file
-p / --prompt	conf/summarize_prompt.tmpl	Summarize prompt template (when STAMPER_SUMMARIZE=true)

---

Keybindings

Key	Action
p	pass
r	review
f	fail
s	summarize via LLM (requires STAMPER_SUMMARIZE=true)
x	skip — conversation is not written to output
Ctrl+C	quit — all completed annotations are preserved
↑ / ↓	scroll turns (TUI mode)

---

Resume

Interrupt at any time with Ctrl+C. Re-run the same command — stamper reads the output file on startup and silently skips already-annotated conversations.

Choosing a different output file starts fresh.

---

Configuration

Copy .env.example to .env and fill in the values. A .env file is optional — any variable can also be exported directly in the shell.

cp .env.example .env

Variable	Default	Description
LLM_FAMILY	openai_platform	LLM provider: openai_platform, openai (Azure), ollama
MODEL_ID	—	Model name, e.g. gpt-4o-mini, llama3
OPEN_AI_KEY	—	OpenAI or Azure API key
AZURE_OPENAI_ENDPOINT	—	Azure OpenAI endpoint URL
OLLAMA_BASE_URL	http://localhost:11434/v1	Ollama base URL
MODEL_MAX_TOKENS	1000	Max tokens for summarize calls
MODEL_TEMPERATURE	0.0	Sampling temperature
STAMPER_SUMMARIZE	false	Enable [s] summarize — no LLM client is created when false
STAMPER_TUI	true	Set to false for plain terminal mode (no bubbletea)

Prompt template

Summarize reads conf/summarize_prompt.tmpl from disk (path relative to the working directory). Ship this file with the binary in release archives; if the file is missing, stamper fails when summarize is enabled. Override with -p:

stamper -i sampled.jsonl -p /path/to/custom_prompt.tmpl

The template has access to .Turns (array of Query / Answer) and the inc helper to produce 1-based turn numbers.

---

Input format

JSONL — one conversation per line:

{"conversation_id": "conv-001", "turns": [{"user_query": "What is Python?", "answer": "A high-level language..."}]}
{"conversation_id": "conv-002", "turns": [{"user_query": "Is it hard?", "answer": "Not at all..."}]}

Field	Required	Description
conversation_id	yes	Unique identifier
turns	yes	Array of turns
turns[].query	yes	User message
turns[].answer	yes	Agent response
(anything else)	no	Preserved as-is in output

Output format

Original JSON with human_annotation appended:

{"conversation_id": "conv-001", "turns": [...], "human_annotation": "pass"}

Skipped conversations (x) produce no output line.

---

Development

# Run tests
go test ./...

# Lint
golangci-lint run

# Sample dataset (20 conversations)
stamper -i resources/sampled.jsonl