Datasets:

Heisenger
/

sft-benchmark-trajectories

code: https://github.com/heisenger/nvidia_benchmark, tag v0.3.0;
dataset: https://huggingface.co/datasets/Heisenger/sft-benchmark-trajectories;
immutable dataset revision: 36c9f25415617c20d7367e58c8ed059bb06dfcbe;
canonical manifest SHA-256: 11358b302b25481d6315cfa08eac1bd95c663946d294def383153b2b15da3a73.

The Hugging Face repository contains 76 files occupying 9,536,452,678 bytes. Its public metadata inventory covers all 155,278 canonical trajectories and 67,610 task keys. Canonical event payloads are hosted for 136,373 trajectories. The other 18,905 records retain scores, ranks, lineage, hashes and pinned upstream pointers but are metadata-only:

agentsuite_tau2: 8,340;
browsecomp_rollouts: 5,063;
swegym_sampled: 5,502.

These records were not discarded. Their transformed-redistribution permission remains uncertain, so their raw events are not in the public release. registry/permissions.json separately records public access, local research use, model training, commercial training, original redistribution and transformed redistribution, including verification state, conditions and evidence URLs for every source. This is an engineering audit rather than legal advice.

Install and download

Python 3.11 or newer is required.

python3 -m venv .venv
.venv/bin/pip install \
  'git+https://github.com/heisenger/nvidia_benchmark.git@v0.3.0'
.venv/bin/pip install huggingface_hub

.venv/bin/hf download \
  Heisenger/sft-benchmark-trajectories \
  --repo-type dataset \
  --revision 36c9f25415617c20d7367e58c8ed059bb06dfcbe \
  --local-dir ./sft-benchmark-trajectories

Normal training reads this persisted snapshot and does not redownload upstream sources. Initial ingestion and explicit refresh remain separate source-adapter commands documented in trajectory_sources.md.

Consume a policy-selected subset

from agent_trajectories import (
    CanonicalDatasetAdapter,
    DatasetQuery,
    TrainingView,
)

dataset = CanonicalDatasetAdapter("./sft-benchmark-trajectories")
query = DatasetQuery.for_training(
    splits=("train",),
    content_availability=("hosted",),
    training_permissions=("permitted", "permitted_with_conditions"),
    training_verifications=("verified",),
)

for example in dataset.iter_examples(
    query, view=TrainingView.FULL_TRAJECTORY
):
    row = example.to_dict()
    print(row["sample_id"], row["metadata"]["lineage"])
    break

That train-split query returns 77,168 trajectories in this release. It combines the persisted structural/training-quality gates with verified training permission; projects should additionally filter commercial_training_permissions when applicable.

Use TrainingView.ASSISTANT_TURN for causal turn samples. Selection, deterministic sampling, alternatives, batching, evaluation bundles and optional PyTorch integration are documented in dataset_adapter.md.

If a consumer is authorized to retrieve a metadata-only source, first run its explicit pinned source adapter and then federate that local content:

dataset = CanonicalDatasetAdapter(
    "./sft-benchmark-trajectories",
    content_overlays={
        "agentsuite_tau2": "./authorized/agentsuite_tau2-canonical"
    },
)

The overlay is joined by canonical IDs and is never uploaded or copied by the consumer. Without one, a payload request for a metadata-only record raises MissingTrajectoryContentError while metadata and permission queries still work.

Real Nemotron SFT demonstration

The real successful official_tau3 trajectory trajectory:f9cf8588a314fa9804d57402 has 13 events and six assistant targets. Using the tokenizer and native chat template from nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 at revision a9904d24bcc1d289a1950fa9d2b978c47cf903b9 produced:

View	Tokens	Supervised assistant tokens	Ignored tokens
Full trajectory	3,174	696	2,478
One assistant turn	1,877	60	1,817

The turn example masks 80 prior assistant tokens, and both examples use -100 for every non-target label. The demonstrated right-padded batch has shape [2, 3174], fields input_ids, attention_mask and labels, and SHA-256 d680ae7e4ea605b3db96370de6760d86ed227411dac901d74a511f72c3346ae8. See nemotron_sft_demo.md and the published examples/nemotron_sft_demo.json.

Scores and validation

No data was rebuilt or rescored for publication. The release preserves:

upstream native rewards and source-reported success/failure labels;
computed structural completeness, action/observation, known-outcome, verification-reliability and SFT-eligibility scores;
deterministic within-task quality ranks.

It contains zero independent environment reproductions and zero new native benchmark replays. Model NLL has not been computed; the Nemotron result is a tokenization and loss-mask demonstration, not model evaluation.

A fresh virtual environment installed package 0.3.0 from the public Git commit, downloaded the complete dataset revision anonymously, instantiated consumer API 1.1, and verified:

all 155,278 metadata records and all 136,373 hosted payload declarations;
96,369 verified-permission, persisted-gate training records;
portable manifest paths with no original workspace references;
one full and six causal assistant-turn views for the real trajectory above;
the metadata-only payload guard and all 13 queryable source permission audits.

The primary limitations are source-reported rather than replayed outcomes, source-specific use conditions, text-encoded tools in some corpora, and metadata-only payloads where transformed redistribution remains unresolved.