← LeaderboardThe Skill Leaderboard

make-data-pipeline

Port an open-source bioinformatics data pipeline into the BioIR pipeline architecture. Use when the user asks to create, port, convert, or write a new data pipeline from OSS code, or add a new model's data processing to the BioIR system. Level 2 is one full build_processor run of the sample manifest, serial and Ray.

Published by NVIDIA ↗NVIDIA/skills

Try your first task

A source-grounded starter task has not been verified for this skill yet.

SKILL.md

# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.# SPDX-License-Identifier: Apache-2.0## Licensed under the Apache License, Version 2.0 (the "License");# you may not use this file except in compliance with the License.# You may obtain a copy of the License at## http://www.apache.org/licenses/LICENSE-2.0## Unless required by applicable law or agreed to in writing, software# distributed under the License is distributed on an "AS IS" BASIS,# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.# See the License for the specific language governing permissions and# limitations under the License.name: make-data-pipelinedescription: Port an open-source bioinformatics data pipeline into the BioIR pipeline architecture. Use when the user asks to create, port, convert, or write a new data pipeline from OSS code, or add a new model's data processing to the BioIR system. Level 2 is one full build_processor run of the sample manifest, serial and Ray.license: Apache-2.0metadata:   author: Tien Pham

Port an Open-Source Data Pipeline to BioIR

Input: the open-source model's name and source location. Output: a registered BioIR pipeline, validation evidence, and a summary report.

The procedure lives in the BioIR checkout, written $BIOIR below. Without a checkout, clone https://github.com/NVIDIA-BioNeMo/BioNeMo-Inference-Runtime and work on a branch. Read both guides in full before you start. Then follow the port guide's steps in order, and use the validation guide for the Level 1 and Level 2 checks. Start each step only after the step before it passes. When a guide step branches, take the branch its stated condition selects, and record that choice in implementation-notes.md:

  1. $BIOIR/docs/advanced/port-data-pipeline.md
  2. $BIOIR/docs/advanced/validate-data-pipeline.md

This file lists only what an agent must do on top of them. Level 2 is one full build_processor run of the manifest, defined below. The port guide's examples/folding/run_demo.py command folds one bundled file and only shows that the stages connect. That command does not satisfy Level 2.

Instructions

Answer an offline audit from the recorded input. Do not run a pipeline, clone a repository, or use the network for that request. A port follows the steps below.

  1. Gather every input before you start: the upstream source location and revision, the WORKDIR (default workdir/<model_name>/), the checkpoint, the samples, and the scorers. If more than one source revision is visible and the user has not named one, ask the user. Do not guess.
  2. Follow port-data-pipeline.md, then validate-data-pipeline.md. Start each step only after the previous one passes.
  3. Run Level 1 on the whole sample manifest. Record the shared settings in NOTES.md.
  4. Run Level 2, the full pipeline on serial and Ray, only after Level 1 passes and the settings are recorded. Copy the seed and runtime_args from NOTES.md into run_pipeline.py. Never copy example numbers from the guide.
  5. Score every structure with the OpenStructure and DockQ commands. Apply the 0.05 rule and the INVESTIGATE rule from the validation guide.
  6. Log each decision in implementation-notes.md, then write the final report.

Examples

  • Audit a Level 2 run. Given a recorded run and its manifest IDs, compare the processed IDs with the manifest. Missing, extra, or duplicate IDs mean the run does not satisfy Level 2.
  • A demo is not Level 2. A passing run_demo.py on one bundled file shows only that the stages connect. It does not satisfy Level 2.
  • A score above upstream. A BioIR lDDT above upstream, with matching serial and Ray scores, is INVESTIGATE. It stays INVESTIGATE until the validation guide's Check Feature Equivalence section passes, NOTES.md records identical settings for both runs, and one scoring command produced both scores.
  • Requests for other skills. A module swap goes to accelerate-custom-model. A folding benchmark goes to bench-perf-oss. Neither starts a data pipeline port.
  • An ambiguous revision. Two source revisions are visible and the user has not named one, and the hub does not resolve a checkpoint. Ask the user which revision to use, and do not start the port.

Full pipeline run

Level 1 compares features and stops before the engine and the writer. Level 2 is the full data pipeline: parser, tokenizer, feature generator, engine, and writer, on every id in ref_data/sample_manifest.json. Run that set once through the serial processor and once through Ray. One request, or one file under examples/data/samples/, is not this run.

Start Level 2 only after Level 1 passes on the whole manifest, and only after ref_data/upstream_config.json and NOTES.md record the shared settings. runtime_args and the seed come from NOTES.md. Copy the numbers in the validation guide only when NOTES.md records those numbers.

The script below uses the folding example's load_requests and the CIF writer. For a non-folding model, replace load_requests with the model's own loader, and use the output format the port guide names. Record that change in implementation-notes.md. If the output cannot be scored with OpenStructure or DockQ, score it with the metric the validation guide names for that output type, and record that substitution. Do not report an OpenStructure or DockQ result for it.

Write $WORKDIR/run_pipeline.py and execute it. The validation guide's snippet leaves REPO, WORKDIR, load_requests, and assert_manifest_coverage to earlier sections, so the script has to define them. The script does not parse NOTES.md. Copy the seed, runtime_args, and any modified configuration into SEED, RUNTIME_ARGS, and ENGINE_KWARGS by hand. The script refuses to run while any of those three is None. If NOTES.md does not record one of them, stop and report which value is missing. Do not guess a seed or a configuration. Inside the development container, drop the uv run --locked prefix.

export WORKDIR="$HOME/workdir/<model_name>"
uv run --locked python "$WORKDIR/run_pipeline.py"
import json
import os
import sys
from pathlib import Path

import ray

import bionemo_ir
from bionemo_ir.pipeline.processor.engine_proc import EngineProcessorConfig, build_processor
from bionemo_ir.pipeline.stages.configs import FeatureGeneratorStageConfig, WriterStageConfig

REPO = Path(bionemo_ir.__file__).resolve().parents[1]
WORKDIR = Path(os.environ["WORKDIR"])
sys.path.insert(0, str(REPO / "examples" / "folding"))
from run_demo import load_requests  # noqa: E402

MODEL_NAME = "<registered model name>"
SEED = None  # the seed recorded in NOTES.md
RUNTIME_ARGS = None  # the runtime_args recorded in NOTES.md
ENGINE_KWARGS = None  # {} or {"config": config} recorded for Level 1


def assert_manifest_coverage(processed_ids):
    manifest_path = WORKDIR / "ref_data" / "sample_manifest.json"
    with open(manifest_path) as handle:
        expected = {sample["sample_id"] for sample in json.load(handle)["samples"]}
    processed = list(processed_ids)
    if len(processed) != len(set(processed)):
        raise SystemExit("duplicate sample IDs")
    missing, extra = expected - set(processed), set(processed) - expected
    if missing or extra:
        raise SystemExit(f"missing={sorted(missing)} extra={sorted(extra)}")


def load_rows():
    with open(WORKDIR / "ref_data" / "sample_manifest.json") as handle:
        manifest = json.load(handle)
    rows = []
    for sample in manifest["samples"]:
        matched = [
            request
            for request in load_requests(REPO / sample["input"])
            if request["input_id"] == sample["sample_id"]
        ]
        if not matched:
            raise SystemExit(f"no request matches manifest id {sample['sample_id']}")
        rows.extend({"record": r, "__record_id": r["input_id"], "random_seed": SEED} for r in matched)
    return rows


def processor_config(executor_backend, output_path):
    return EngineProcessorConfig(
        model_source=MODEL_NAME,
        executor_backend=executor_backend,
        engine_kwargs=ENGINE_KWARGS,
        runtime_args=RUNTIME_ARGS,
        feature_generator_stage=FeatureGeneratorStageConfig(init_context={"random_seed": SEED}),
        writer_stage=WriterStageConfig(output_path=output_path, format="cif"),
    )


def run(executor_backend, output_path):
    if output_path.exists() and any(output_path.iterdir()):
        raise SystemExit(f"output directory is not empty: {output_path}")
    rows = load_rows()
    processor = build_processor(processor_config(executor_backend, str(output_path)))
    if executor_backend == "ray":
        results = list(processor(ray.data.from_items(rows)).materialize().iter_rows())
    else:
        results = processor(rows)
    failed = [row["__record_id"] for row in results if (row.get("__inference_error__") or {}).get("error_msg")]
    if failed:
        raise SystemExit(f"inference failed: {failed}")
    assert_manifest_coverage(row["__record_id"] for row in results)
    print(f"{executor_backend or 'serial'}: {len(results)} structures in {output_path}")
    del processor, results


def release_serial_cuda():
    import gc

    import torch

    gc.collect()
    if torch.cuda.is_available():
        torch.cuda.empty_cache()


if __name__ == "__main__":
    if SEED is None or RUNTIME_ARGS is None or ENGINE_KWARGS is None:
        raise SystemExit("set SEED, RUNTIME_ARGS, and ENGINE_KWARGS before running")
    run(None, WORKDIR / "bioir" / "serial")
    # The serial processor ran in this process. Return its cached CUDA blocks
    # before Ray workers allocate.
    release_serial_cuda()
    run("ray", WORKDIR / "bioir" / "ray")

The two runs share the driver process. After the serial run returns, the script drops that processor and its outputs, collects garbage, and calls torch.cuda.empty_cache(). PyTorch otherwise keeps the unused CUDA blocks in the driver cache, and Ray workers cannot allocate that memory.

If Level 1 used a modified configuration, set ENGINE_KWARGS = {"config": config} in the script. Both runs read that one object, so the serial and Ray runs cannot diverge.

A passing script exits 0, prints both lines, and leaves one structure per manifest id under bioir/serial and bioir/ray. The script stops when either directory already holds files, so a stale structure cannot pass coverage. Remove that directory before a rerun. Cite the command, the exit code, both directories, and the sample ids from the manifest. Then score every structure with the OpenStructure and DockQ commands in the validation guide, and apply that guide's 0.05 rule. Serial and Ray scores for a sample must match. A BioIR score above upstream is INVESTIGATE, not PASS, until feature equivalence, identical settings, and the same scoring command are confirmed.

Hard Gates

Treat the rules for trustworthy results in $BIOIR/docs/advanced/validate-data-pipeline.md as hard gates. In addition:

  • Every PASS you report cites the command, its exit code, the artifact path, and the exact sample set. Never report a result from memory or partial output.
  • A gate you cannot meet is a blocker. Stop and report it, and do not claim completion.
  • Changing a frozen tolerance, threshold, or feature category, or narrowing the sample scope, needs user approval and a rationale in NOTES.md.

Ask the User

  • First: where the upstream source lives, and the WORKDIR (default workdir/<model_name>/). Do not guess which version or fork is authoritative.
  • Before: extending an existing pipeline instead of creating one, adding a dependency, or downloading a checkpoint that bionemo_ir.hubs does not resolve.
  • Immediately: when you find an upstream bug.

The user might step away. Gather every resource — source, checkpoint, samples, scorers — at the start.

Working Directory

Set up the environment as in $BIOIR/docs/dev.md, and the WORKDIR as in $BIOIR/docs/advanced/validate-data-pipeline.md. Never commit the WORKDIR.

Keep $WORKDIR/implementation-notes.md as a running log, written when a decision happens rather than at the end. Each entry has a timestamp, a category (design decision, deviation, tradeoff, or open question), the context, the decision, the reason, alternatives considered, the expected impact, and whether it needs user confirmation.

Final Report

Consolidate implementation-notes.md so a future agent can resume from it, then print a summary:

  1. Model, upstream location and commit, and pipeline shape.
  2. The function inventory: upstream function to BioIR class. Record each mapping in implementation-notes.md as you make it, so this report only consolidates it.
  3. Files created or modified.
  4. Level 1 results, then the full pipeline run: command, exit code, serial and Ray artifact directories, and the manifest ids. Include the OpenStructure and DockQ comparison.
  5. Limitations, and anything that needs human review.
  6. The WORKDIR, the path to implementation-notes.md, and every open question.