Datasetpublic

forseasons/dab-fsm-grounded-rl-300-20260725

DAB FSM-Grounded Altimate RL 300 This public package contains 300 replay-grounded synthetic DataAgentBench-style tasks compiled for the dab_sandbox_altimate_noctx VERL/Altimate runtime. Every task passed: FSM reachability and required-action execution checks real database evidence replay unique-answer materialization strict validator positive and negative self-tests RL LLM quality judging DAB-style query surface realization final sandbox task materialization checks… See the full description on the dataset page: https://huggingface.co/datasets/forseasons/dab-fsm-grounded-rl-300-20260725.

sourceHugging Faceupdated 2mo agoView on Hugging Face

0likes53downloads

Dataset Card

DAB FSM-Grounded Altimate RL 300

This public package contains 300 replay-grounded synthetic DataAgentBench-style tasks compiled for the dab_sandbox_altimate_noctx VERL/Altimate runtime.

Every task passed:

—FSM reachability and required-action execution checks
—real database evidence replay
—unique-answer materialization
—strict validator positive and negative self-tests
—RL LLM quality judging
—DAB-style query surface realization
—final sandbox task materialization checks

Main files

File	Description
`train.parquet`	300 VERL-ready RL rows
`test.parquet`	Compatibility mirror of the same 300 rows
`train.jsonl`, `test.jsonl`	JSONL mirrors
`validators.jsonl`	All 300 verifier specs and complete `validate.py` sources
`task_artifacts.tar.gz`	Complete task artifacts, including validators, evidence graphs, replay traces, blueprints, descriptions and DB configs
`sandbox_task_manifest.json`	Runtime task manifest
`task_artifacts_manifest.json`	Full artifact index and provenance
`task_list.json`	Compact task list
`scripts/materialize_dabench_assets.py`	Materialization helper

The train and test files are mirrored runtime-compatible exports, not an independent benchmark split.

Validator behavior

The reward path remains:

text

sandbox task artifact
  -> query*/validate.py
  -> strict semantic comparison
  -> binary DAB reward

Canonical JSON remains accepted. For DAB-like natural questions that do not explicitly request JSON, flat object/list validators also accept a compact one-line or line-oriented answer containing exactly the required values. Wrong values, wrong ordering, duplicates, prose and additional payload remain rejected.

validators.jsonl makes every validator independently inspectable:

json

{
  "dataset": "query_synthetic_...",
  "query_id": 1,
  "artifact_path": "task_artifacts/query_synthetic_.../query1",
  "query": {},
  "verifier_spec": {},
  "validate_py_sha256": "...",
  "validate_py": "..."
}

Extract task artifacts

bash

tar -xzf task_artifacts.tar.gz

The archive contains 300 validate.py files and their corresponding verifier_spec.json, query.json, task_blueprint.json, task_evidence_graph.json, fsm_execution_replay.json, replay trace and difficulty calibration.

Runtime assets

This is a delta package against an existing DataAgentBench sandbox. It does not redistribute the upstream database dumps. After downloading the corresponding DAB assets, materialize the task directories with:

bash

python scripts/materialize_dabench_assets.py \
  --package-dir /path/to/this_dataset \
  --bench-root /mnt/docker-data/workspace/DataAgentBench \
  --canonical-bench-root /mnt/docker-data/workspace/DataAgentBench \
  --overwrite

Mount the resulting benchmark root at /workspace/DataAgentBench in the sandbox and use train.parquet or test.parquet with the dab_sandbox_altimate_noctx data source.

Runtime contract

text

data_source=dab_sandbox_altimate_noctx
DAB_AGENT_LOOP_VARIANT=altimate_noctx
DAB_AUTO_CONTEXT_MODE=none
DAB_MAX_ITERATIONS=200
response_length=49152

Pipeline code: sunnweiwei/DataAgent@0725_v1

Use of the underlying database assets remains subject to the upstream DataAgentBench dataset terms.