teich 0.2.1

Turn coding agent traces into auditable supervised fine-tuning data

Teich

Agent data infrastructure for generation, normalization, formatting, response masking, and training audits.

Teich turns raw agent sessions, chat datasets, local JSONL, Hugging Face datasets, and in-memory datasets.Dataset objects into auditable SFT data.

It handles the parts that usually break training runs:

• normalizing traces into OpenAI-style messages and tools
• preserving tool schemas, reasoning, metadata, and provenance
• rendering through your target tokenizer's chat template
• recording typed supervision spans before tokenization
• applying response-only labels after TRL / Unsloth trainer tokenization
• reporting dropped, oversized, trimmed, malformed, and fully masked rows

Use it as a trace generator, a dataset loader, a chat-template renderer, a masking layer, or the whole pipeline.

Install

pip install teich

Or run it without installing:

uvx teich --help

Agent trace generation needs Docker and an API key for the configured provider. Preparing an existing local or Hugging Face dataset does not need Docker.

Prefer a browser workflow?

teich studio

See Teich Studio.

Quickstart: Prepare Existing Data

If your dataset already has messages, Teich can usually prepare it directly.

from teich import prepare_data

train_dataset = prepare_data(
    "TeichAI/Claude-Opus-4.6-Reasoning-887x",
    tokenizer,
    max_length=32768,
    oversized_policy="trim_followups",
    tokenize=True,
    chat_template_kwargs={"enable_thinking": True, "preserve_thinking": True},
)

Then create your trainer and call mask_data():

from teich import mask_data

trainer = mask_data(
    trainer,
    tokenizer=tokenizer,
    train_on_reasoning=True,
    train_on_final_answers=True,
    train_on_tools=True,
)

More detail: Preparing Data and Training.

Quickstart: Generate New Traces

teich init my-project
cd my-project

Add prompts to prompts.jsonl:

{"prompt":"Build a simple todo list app in React"}
{"github_repo":"armand0e/perplexica-mcp","prompt":"Add a small usability improvement and update the tests"}
{"prompt":"Draft a compact project plan","follow_up_prompts":["Revise it for a solo developer","Add a risk checklist"]}

Set your provider key and run:

export OPENAI_API_KEY=sk-...
teich generate -c config.yaml

Teich writes raw traces, converted training rows, sandbox snapshots, and a dataset card under output/. Use --resume to skip prompts that already completed.

More detail: Generation.

Quickstart: Extract Local Sessions

If you already have local agent sessions, Teich can stage them as an anonymized dataset in one command:

teich extract claude --model fable-5

extract supports claude, codex, pi, and hermes. It writes anonymized traces to data/ by default, with JSONL files directly in that folder so the generated Hugging Face dataset metadata can match *.jsonl. It generates a dataset README.md, and then asks whether to upload the folder to Hugging Face. Use --out / --output to choose another folder.

What Teich Supports

Use case	Start here
Configure and steer runs in a browser	Teich Studio
Generate Codex, Pi, Claude Code, Hermes, or chat data	Generation
Load local files, folders, Hugging Face datasets, or datasets.Dataset objects	Preparing Data
Train with TRL / Unsloth while keeping response-only labels correct	Training
Understand messages, tools, metadata, and native trace behavior	Data Format
Use prepare_data, mask_data, load_traces, and validation helpers	Python API
See the full generation, preparation, and masking pipeline	Pipeline Flow

Why Teich

Most SFT pipelines flatten agent data too early. That loses tool schemas, tool results, reasoning boundaries, provenance, and the exact assistant spans you meant to train on.

Teich keeps the data structured until the last practical moment:

prompts / traces / JSONL / HF datasets / Dataset objects
        -> load_traces() or prepare_data()
        -> normalized messages + tools
        -> tokenizer chat template rendering
        -> trainer-friendly text + Teich supervision spans
        -> SFTTrainer tokenization
        -> mask_data()
        -> audited input_ids + labels

This makes multi-turn, tool-call, reasoning, and mixed-source datasets trainable without relying on brittle single-span masking.

Common Commands

# Create a generation project
teich init my-project

# Generate data from config.yaml
teich generate -c config.yaml

# Resume an interrupted batch
teich generate -c config.yaml --resume

# Extract, anonymize, and stage local Claude Code traces
teich extract claude --model fable-5 --out data

# Launch the local browser UI
teich studio

# Use a local OpenAI-compatible endpoint
TEICH_PROVIDER=LMstudio \
TEICH_MODEL=gemma-4 \
TEICH_BASE_URL=http://localhost:1234/v1 \
TEICH_API_KEY=llm \
teich generate -c config.yaml

Minimal Config

agent:
  provider: codex  # codex, pi, claude-code, hermes, or chat

model:
  model: codex-mini-latest
  approval_policy: never
  sandbox: danger-full-access

prompts_file: prompts.jsonl

output:
  traces_dir: ./output
  sandbox_dir: ./sandbox
  failures_dir: ./failures

publish:
  repo_id: username/my-dataset
  private: false

agent.provider: chat writes structured chat rows directly and does not require Docker. Agent providers preserve raw or native traces as source-of-truth artifacts.

Python Entry Points

from teich import (
    prepare_data,
    mask_data,
    load_traces,
    detect_trace_type,
    validate_tool_calls,
    row_fits_context,
    trace_is_complete,
    preview_sft_example,
)

See Python API for the full public surface.

Status

Teich is alpha. The core trace, preparation, masking, and audit workflow is usable, but APIs may evolve as more agent formats and training flows are added.

Development

uv pip install -e ".[dev]"
uv run pytest --ignore=tests/test_integration.py -q

License

Apache-2.0