Heisenger/sft-benchmark-trajectories

SFT Benchmark Trajectories

A model-agnostic collection of real expert demonstrations and recorded agent completions, normalized into one interface for full-trajectory and causal assistant-turn SFT.

At a glance

• 260,753 canonical records across 154,244 task keys
• 13 benchmark families and 21 trajectory sources
• 235,089 hosted payloads plus 25,664 metadata-only records
• 67,017 success-labelled, 88,350 failure-labelled, and 105,386 unknown-outcome records
• Task-disjoint train, validation, test and protected-evaluation splits
• Implementation version: v0.5.0 (installable wheel included)

The inventory includes full recorded trajectories and explicitly marked assistant-turn prefixes. Prefix records provide exactly one supervised answer with its published context; they are never presented as complete episodes.

Priority SFT selections

Source	Available selection	Training views	Outcome caveat
AReaL τ² SFT	areal_tau2_author_sft: all 33,531 original targets	1,824 reconstructed full dialogs; all retained records support turns	correct and reward disagree on 2,078 original rows; both are preserved
LiteCoder Terminal SFT	litecoder_terminal_author_sft: 11,255 original conversations	Full recorded interaction and assistant turns	No per-record grader score; self-claimed completion is not verified success
ToolBench / ToolLLaMA	toolbench_valid_solution_paths: 73,121 existing paths	Full and turn	Official valid-data selection; semantic success unavailable
SWE-smith	swesmith_sweagent_lm32b_author: 3,567 existing successful records	Full and turn	Exact intersection with 4,954 author-published IDs; 1,387 are unavailable in our tool serialization
SWE-Gym OpenHands SFT	swegym_openhands_author_sft: all 491 existing successful records	Full and turn	Published success-only split; no local replay

These are selections over the same canonical records. Selecting them adds no duplicate trajectory copies. The SFT expansion audit documents the exact source releases, preserved target counts and limitations.

Historical outcomes below are the results attached by each source. This project did not rerun the benchmark environments. A source score describes the imported attempts, not the expected score of a newly trained model.

Benchmarks

Benchmark	What it tests	Native or published metric	Tasks	Records
BrowseComp	Research-style web browsing and answer synthesis.	answer correctness (llm judge)	1,266	5,063
Mind2Web	Human demonstrations for completing tasks on real websites.	step and element accuracy (action match)	1,009	1,009
SWE-bench	Resolve real software repository issues.	resolved (native binary)	34,244	56,898
SWE-Gym	Train and evaluate software-engineering agents on repository tasks.	execution-test success (native binary)	2,731	5,993
SWE-smith	Generated software-engineering tasks with executable validation.	resolved (native binary)	10,637	24,100
τ2-bench	Tool-using agents interacting with simulated users in service domains.	evaluator correctness and reward (native reward)	2,727	21,962
τ3-bench	Tool-using agents and simulated users with environment reward evidence.	environment/action/communication reward (native reward)	50	200
Terminal-Bench	Complete practical tasks in sandboxed terminal environments.	task reward (native reward)	10,161	45,518
TerminalTraj	A corpus of terminal command and observation trajectories.	not available (none)	17,787	20,000
ToolBench	Tool-use planning and API-call trajectories.	valid-data/win generation status (source proxy)	73,132	73,132
BFCL V4	Agentic function-calling tasks with web and memory tool environments.	BFCL evaluator score (native binary)	665	6,160
DeepSWE	Repository engineering tasks executed by multi-agent DeLM teams.	CTRF test-suite pass (native binary)	10	360
Terminal-Bench 4	Protected terminal tasks executed by multi-agent DeLM teams.	task reward (native reward)	10	358

Benchmark rows are not additive: 185 task keys appear in more than one benchmark family.

Metrics are deliberately not combined into one global score. Binary repository tests, scalar rewards, LLM judges and unscored human demonstrations are not interchangeable.

Trajectory sources and observed performance

Benchmark	Trace source and origin	Source ID	Tasks	Records	Score in imported slice	Score context	Payload
BrowseComp	BrowseComp GPT-OSS-120B Rollouts	browsecomp_rollouts	1,266	5,063	23.92% success	5,063 scored attempts; Source-reported GPT-4o answer judge; not independently reproduced	metadata_only
Mind2Web	Mind2Web	mind2web	1,009	1,009	not available	task outcome unavailable; Per-action DOM context and human annotation are present; no explicit binary execution-success field	hosted
SWE-bench	NVIDIA Open-SWE-Traces	nvidia_open_swe_traces	33,444	51,626	47.93% success	51,626 scored attempts; Published resolved field; not independently replayed here	hosted
SWE-bench	SWE-bench Multilingual GLM-5 Trajectories	swebench_multilingual_glm5	300	300	69.67% success	300 scored attempts; Published resolved field	metadata_only
SWE-bench	SWE-bench Multilingual MiniMax-M2.5 Trajectories	swebench_multilingual_minimax_m25	299	299	73.58% success	299 scored attempts; Published resolved field	metadata_only
SWE-bench	SWE-bench Multilingual Trajectories	swebench_multilingual_trajectories	43	129	6.20% success	129 scored attempts; Published resolved/completed/errored fields	hosted
SWE-bench	SWE-bench Verified mini-swe-agent Trajectories	swebench_verified_trajectories	500	4,544	74.65% success	4,544 scored attempts; Published SWE-bench execution result; protected evaluation split	hosted
SWE-Gym	SWE-Gym OpenHands SFT Trajectories	swegym_openhands	293	491	100.00% success	success-only release; Published train.success.oss split label; not independently replayed	hosted
SWE-Gym	SWE-Gym OpenHands Sampled Trajectories	swegym_sampled	2,438	5,502	8.92% success	5,502 scored attempts; Source execution test result; not independently replayed here	metadata_only
SWE-smith	SWE-smith Trajectories	swesmith_trajectories	10,637	24,100	39.12% success	24,100 scored attempts; Published resolved field	hosted
τ2-bench	AgentSuite tau2-bench trajectories	agentsuite_tau2	228	8,340	0.478 mean reward; 40.47% success	8,340 scored attempts; AgentSuite evaluator fields; meta.is_correct is used for the binary outcome	metadata_only
τ2-bench	AReaL tau2 SFT	areal_tau2_sft	2,499	13,622	0.798 mean reward; 74.46% correct labels	33,531 targets; published dialog labels, weighted over original supervised targets; Source correctness/reward labels, with 2,078 turn rows whose labels disagree; no grader or independent replay	hosted
τ3-bench	Official tau3-bench	official_tau3	50	200	0.500 mean reward; 50.00% success	200 scored attempts; Environment/action/communication reward evidence	hosted
Terminal-Bench	Terminal-Bench 2.0 Recorded Trajectories	terminalbench_trajectories	89	34,263	0.323 mean reward; 32.25% success	34,263 scored attempts; Source-reported Terminal-Bench reward	hosted
Terminal-Bench	LiteCoder Terminal SFT	litecoder_terminal_sft	10,072	11,255	not available	task outcome unavailable; Original SFT selection; agent task_complete claims are not grader evidence	hosted
TerminalTraj	TerminalTraj	terminaltraj	17,787	20,000	not available	task outcome unavailable; Paired command/output transcript; no trustworthy task-level success field	hosted
ToolBench	ToolBench	toolbench	11	11	63.64% success	11 scored attempts; valid_data/win generation status; no independent semantic proof	hosted
ToolBench	ToolBench Full Valid Solution Paths	toolbench_full	73,121	73,121	not available	task outcome unavailable; Official preprocessing includes only valid_data paths; search-win/task-success labels are absent	hosted
BFCL V4	AgentSuite BFCL V4 Agentic Trajectories	agentsuite_bfcl_v4	665	6,160	0.784 mean reward; 78.43% success	6,160 scored attempts; Published BFCL V4 evaluator score	metadata_only
DeepSWE	DeLM DeepSWE Multi-Agent Trajectories	delm_deepswe	10	360	0.917 mean reward; 91.67% success	360 scored attempts; Published DeepSWE CTRF test report	hosted
Terminal-Bench 4	DeLM Terminal-Bench 4 Multi-Agent Trajectories	delm_terminalbench4	10	358	0.863 mean reward; 86.27% success	357 scored attempts; Published Terminal-Bench 4 grader reward	hosted

success-only release means the dataset was conditioned on a successful result; it does not imply that the generating agent had a 100% policy success rate. metadata_only sources retain scores, lineage, hashes and pinned source pointers but require a consumer-authorized content overlay. Aligned synthetic SFT sources contain their own training tasks. Their task counts do not represent the size of the official benchmark evaluation set.

Model context

The following rows contextualize trace strength where one source contains multiple sufficiently covered models. Models are scored on the intersection of their task keys within that source. Representative lower, median and higher rows are shown; they are not cross-benchmark rankings.

Benchmark	Source	Model	Position	Common tasks	Scored attempts	Success
BFCL V4	agentsuite_bfcl_v4	openai/gpt-4.1-mini	lower	91	91	82.42%
BFCL V4	agentsuite_bfcl_v4	deepseek-ai/DeepSeek-V3.2-Exp-thinking-off	median	91	91	87.91%
BFCL V4	agentsuite_bfcl_v4	deepseek-ai/DeepSeek-V3.2-Exp-thinking-on	higher	91	91	91.21%
τ2-bench	agentsuite_tau2	togetherai/openai/gpt-oss-20b	lower	228	278	10.79%
τ2-bench	agentsuite_tau2	deepseek-ai/DeepSeek-R1-0528	median	228	278	43.53%
τ2-bench	agentsuite_tau2	anthropic/claude-4.5-sonnet-thinking-on-10k	higher	228	278	54.68%
DeepSWE	delm_deepswe	claude-opus-5-5	lower	10	180	90.56%
DeepSWE	delm_deepswe	gpt-6-astra	higher	10	180	92.78%
Terminal-Bench 4	delm_terminalbench4	gpt-6-astra	lower	10	180	80.56%
Terminal-Bench 4	delm_terminalbench4	claude-opus-5-5	higher	10	177	92.09%
SWE-bench	nvidia_open_swe_traces	Qwen3.8-27B	lower	18,182	18,182	50.71%
SWE-bench	nvidia_open_swe_traces	Qwen3.8-Flash-Next	higher	18,182	18,182	53.60%
SWE-bench	swebench_multilingual_trajectories	Qwen2.5-Coder-32B-Instruct	lower	43	43	0.00%
SWE-bench	swebench_multilingual_trajectories	Multi-SWE-agent-32B-Rust	median	43	43	2.33%
SWE-bench	swebench_multilingual_trajectories	SWE-agent-LM-32B	higher	43	43	16.28%
SWE-bench	swebench_verified_trajectories	gpt-5-mini	lower	44	44	40.91%
SWE-bench	swebench_verified_trajectories	claude-4.5-opus-high	median	44	44	70.45%
SWE-bench	swebench_verified_trajectories	gemini-3-pro	higher	44	44	100.00%
Terminal-Bench	terminalbench_trajectories	openai/gpt-oss-20b@together_ai	lower	80	1,627	3.56%
Terminal-Bench	terminalbench_trajectories	gpt-5.1-codex@openai	median	80	396	39.39%
Terminal-Bench	terminalbench_trajectories	claude-opus-4.6@anthropic	higher	80	153	76.47%

Sources without at least two models covering half of the same source task set are omitted rather than compared on unrelated questions. Full model, scaffold and per-trajectory results remain queryable in index.sqlite and evaluation/summary.json.

No comparable multi-model slice is available here for: BrowseComp, Mind2Web, SWE-Gym, SWE-smith, τ3-bench, TerminalTraj, ToolBench. Those benchmarks have one recorded model, human/unknown provenance, no outcome, or model runs on materially different task slices.

Install and download

python3 -m venv .venv
.venv/bin/pip install 'https://huggingface.co/datasets/Heisenger/sft-benchmark-trajectories/resolve/v0.5.0/code/agent_trajectory_datasets-0.5.0-py3-none-any.whl'
.venv/bin/pip install huggingface_hub
.venv/bin/hf download Heisenger/sft-benchmark-trajectories --repo-type dataset --revision v0.5.0 --local-dir ./sft-benchmark-trajectories

Normal training reads this persisted release. It does not re-download the upstream benchmark datasets. The release includes an installable Python wheel. GitHub publication of the v0.5.0 implementation is pending restored repository authentication; the wheel provides the complete consumer API in the meantime.

Two SFT views from the same record

The consumer selects a view at iteration time. Both views preserve the same task, trajectory, source, split, outcome, reward, verification, quality and permission lineage.

from agent_trajectories import CanonicalDatasetAdapter, DatasetQuery, TrainingView

dataset = CanonicalDatasetAdapter("./sft-benchmark-trajectories")
query = DatasetQuery(trajectory_ids=("trajectory:f9cf8588a314fa9804d57402",))

full = next(dataset.iter_examples(query, view=TrainingView.FULL_TRAJECTORY))
turns = list(dataset.iter_examples(query, view=TrainingView.ASSISTANT_TURN))

Full-trajectory view

One training sample contains the complete causal trace. The event loss mask selects every supervised assistant event.

{
  "view": "full-trajectory",
  "sample_id": "full:736e63752d3cff92f0d883d1",
  "trajectory_id": "trajectory:f9cf8588a314fa9804d57402",
  "task_key": "tau:airline:0",
  "event_count": 13,
  "assistant_targets": 6,
  "loss_mask": [false, true, false, true, false, true, false, true, false, true, false, true, false],
  "metadata": {"source_id": "official_tau3", "success": true}
}

Assistant-turn view

The same trajectory produces one sample per supervised assistant decision. Each sample contains only its causal prefix and one target; prior assistant tokens are context and future observations are absent.

{
  "view": "assistant-turn",
  "sample_id": "turn:61eec58539653d88c1971b68",
  "trajectory_id": "trajectory:f9cf8588a314fa9804d57402",
  "task_key": "tau:airline:0",
  "target_index": 5,
  "context_events": 5,
  "target": {
    "role": "assistant",
    "content": "Let me check the details of your reservation...",
    "tool_call": {
      "name": "get_reservation_details",
      "arguments": {"reservation_id": "EHGLP3"}
    }
  },
  "metadata": {"source_id": "official_tau3", "success": true}
}

Full-trajectory SFT teaches the complete multi-step behavior in one sequence. Turn-based SFT teaches individual decisions conditioned on the state available at that turn. Model-specific tokenization maps the event targets to token-level labels downstream.

Select a training population

from agent_trajectories import (
    CanonicalDatasetAdapter,
    DatasetQuery,
    TrainingView,
    TrajectorySelection,
)

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

for example in dataset.iter_examples(
    query,
    view=TrainingView.FULL_TRAJECTORY,
):
    row = example.to_dict()
    model_input = row["model_input"]
    event_loss_mask = row["loss_mask"]
    lineage = row["metadata"]["lineage"]

Use ALL_SUCCESSFUL instead of BEST_VERIFIED to retain multiple successful strategies per task. Use all attempts for evaluation or preference construction, not as an undifferentiated positive-SFT target.

Author selections and view availability are additional independent filters:

query = DatasetQuery.for_training(
    source_ids=("areal_tau2_sft",),
    curated_selections=("areal_tau2_author_sft",),
    supported_views=("assistant_turn",),
    splits=("train",),
    selection=TrajectorySelection.ALL_SUCCESSFUL,
)
for example in dataset.iter_examples(query, view=TrainingView.ASSISTANT_TURN):
    row = example.to_dict()  # one target; earlier assistant events are context

For LiteCoder, select its author SFT collection with ALL_ATTEMPTS: its grader outcomes are unknown. available_curated_selections() lists pinned selection metadata. for_training always excludes protected-evaluation task assignments.

Benchmark, source and demonstration provenance are independently selectable. For example, this selects successful recorded completions from two exact sources while keeping protected evaluation traces out of a training export:

query = DatasetQuery.for_training(
    benchmarks=("swe-bench", "swe-smith"),
    source_ids=("nvidia_open_swe_traces", "swesmith_trajectories"),
    demonstration_provenance=("recorded_agent_completion",),
    splits=("train",),
    selection=TrajectorySelection.ALL_SUCCESSFUL,
)

Use demonstration_provenance=("expert_or_reference_demonstration",) for expert/reference-only selection. available_sources() lists exact source IDs; metadata(query) can inspect a selection without loading trajectory payloads.

Why the metadata is exposed

Metadata	Training or evaluation purpose
Task and environment keys	Enforce leakage-safe splits and compare compatible attempts
Source, model and scaffold	Filter, balance and audit heterogeneous generators
Outcome, reward and verification	Select positive traces and define the source-performance reference
Quality tier and task-local rank	Choose the best available alternative without treating the score as globally comparable
Event/action counts	Detect short, long, incomplete or looping traces
Token length	Plan truncation, packing, batch size and compute for a named tokenizer
Permissions and payload status	Build a project-specific compliant subset
Revision and record locator	Reproduce the exact upstream record

Token length is tokenizer-specific, not a canonical measure of quality. The published Nemotron example always names the tokenizer and immutable revision.

Documentation

• Consumer API and sampling
• Canonical schema and ranking
• Scoring and evaluation protocol
• Source provenance and ingestion
• v0.4.0 incremental audit
• v0.5.0 SFT expansion
• Licensing, permissions and federated overlays
• Nemotron tokenization and assistant-only masks
• Third-party notices

Limitations

• Source outcomes were not independently replayed.
• Some sources publish no trustworthy task-level outcome.
• Scores from different benchmark metrics or task slices are not directly comparable.
• Some actions and observations are text-encoded rather than typed tool calls.
• training_eligible is a structural/source gate; permission filters remain a separate consumer decision.
• Metadata-only sources require direct retrieval under the consumer's own authorization.
• The audited ALFWorld replayed-expert source remains gated and is not counted in this release because the publishing account has not accepted its access terms; its pinned locator and permission state remain in the registry.