veri-runner

The open-source training runtime for Veri. This is the code that actually executes your training job on a GPU worker: dataset normalization, reward-function invocation, TRL trainer dispatch, and (for harness-in-the-loop RL) token-exact trajectory capture between your agent harness and the in-training policy.

It is published so you can answer "is this bug in my code or in the platform?" by reading the exact code your job ran, instead of asking support. It works well as context for a coding agent: point Claude Code (or any agent) at this repo together with your job's downloaded trajectory or logs and ask it to walk the relevant path.

What's here

Path	What it is
veri_runner/training_runtime.py	Method dispatch (grpo, grpo_harness, sft_text, dpo), dataset format detection/normalization, reward loading, TRL config construction, checkpoint save
veri_runner/trajectory_proxy.py	The per-rollout HTTP proxy between your unmodified agent harness and the policy server: OpenAI and Anthropic (/v1/messages) adapters, token-exact span capture, episode rendering with assistant-only loss masking, and the token-fidelity / template-drift rejection paths
veri_runner/harness_rollout.py	Rollout orchestration for harness-in-the-loop GRPO: per-rollout env contract, sandboxed harness execution, reward scoring, trajectory JSONL output
contracts/worker_config/	Fixture copies of the exact job config the control plane hands the worker
docs/trajectory-format.md	The trajectory / span JSONL schema (what you download from the job page)
examples/	Sanitized example trajectories, including a token-fidelity rejection to debug against
tests/	The unit tests Veri runs against this code in CI

Not here (closed): worker lifecycle (boot, billing callbacks, termination), the control plane, provisioning, and serving.

Debugging token-fidelity rejections

When a rollout is rejected from training ("template drift" or "token ids absent"), the decision was made in veri_runner/trajectory_proxy.py:

1. render_episode requires every span to have token_fidelity: "captured". Spans without server-side token ids are never trained on.
2. Multi-turn episodes must satisfy strict prefix extension: turn k's prompt token ids must extend turn k-1's prompt+completion exactly (verl-style delta tokenization). A mismatch means the chat template rewrote history (thinking models like Qwen3 do this) and raises TemplateDriftError.
3. If stitching fails, the episode falls back to training on the final turn only (metadata.render_fallback = "last_turn"), never on drifted tokens.

To debug a rejected rollout: download the raw trajectory JSONL from the job page, then check the spans against those rules (or hand the file and this repo to a coding agent and ask which rule fired and why). See examples/rejected_trajectory_template_drift.jsonl for a worked example.

Running the tests

uv run --extra dev pytest

The tests run on plain Python (no GPU, no torch): the heavy dependencies are imported lazily per training method and the tests exercise the logic around them.

How changes ship

This repo is vendored into Veri's private monorepo as a git submodule pinned to an exact commit. Merged PRs here do not change production behavior until Veri bumps that pin, runs the full internal suite (including worker-image and live-GPU tests), and deploys a new worker image. So: contributions are welcome, CI here must stay green, and a green PR is necessary but not sufficient for a change to reach production.

Compatibility contract for contributors:

• veri_runner/*.py must keep working when dropped flat into one directory (the worker imports training_runtime, not veri_runner.training_runtime); that is why cross-module imports use the try/except pattern.
• Public function signatures used by the worker (run_training, load_reward_function(s), resolve_training_rows, render_episode, SpanStore, TrajectoryProxy, HarnessRunner) are stable interfaces; breaking them requires a matching worker change on Veri's side, so open an issue first.
• Tests must not require a GPU or network.

License

Apache-2.0. See LICENSE.