The dataset viewer is not available for this dataset.
Error code: ConfigNamesError
Exception: ValueError
Message:
Expected data_files in YAML to be either a string or a list of strings
or a list of dicts with two keys: 'split' and 'path', but got [{'split': 'sft_warmup', 'path': 'train_sft_240.jsonl'}, {'split': 'rl_stage_1', 'path': 'train_grpo_stage1.jsonl'}, {'split': 'rl_stage_2', 'path': 'train_grpo_stage2.jsonl'}, {'split': 'validation', 'path': 'val_train_80.jsonl'}, {'split': 'synthetic_holdout', 'path': 'holdout_synth_400_frozen.jsonl'}, {'split': 'repo-style_holdout', 'path': 'holdout_repo_style_100_frozen.jsonl'}]
Examples of data_files in YAML:
data_files: data.csv
data_files: data/*.png
data_files:
- part0/*
- part1/*
data_files:
- split: train
path: train/*
- split: test
path: test/*
data_files:
- split: train
path:
- train/part1/*
- train/part2/*
- split: test
path: test/*
PS: some symbols like dashes '-' are not allowed in split names
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/dataset/config_names.py", line 67, in compute_config_names_response
config_names = get_dataset_config_names(
path=dataset,
token=hf_token,
)
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 161, in get_dataset_config_names
dataset_module = dataset_module_factory(
path,
...<4 lines>...
**download_kwargs,
)
File "/usr/local/lib/python3.14/site-packages/datasets/load.py", line 1182, in dataset_module_factory
raise e1 from None
File "/usr/local/lib/python3.14/site-packages/datasets/load.py", line 1157, in dataset_module_factory
).get_module()
~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/load.py", line 613, in get_module
metadata_configs = MetadataConfigs.from_dataset_card_data(dataset_card_data)
File "/usr/local/lib/python3.14/site-packages/datasets/utils/metadata.py", line 153, in from_dataset_card_data
cls._raise_if_data_files_field_not_valid(metadata_config)
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/utils/metadata.py", line 100, in _raise_if_data_files_field_not_valid
raise ValueError(yaml_error_message)
ValueError:
Expected data_files in YAML to be either a string or a list of strings
or a list of dicts with two keys: 'split' and 'path', but got [{'split': 'sft_warmup', 'path': 'train_sft_240.jsonl'}, {'split': 'rl_stage_1', 'path': 'train_grpo_stage1.jsonl'}, {'split': 'rl_stage_2', 'path': 'train_grpo_stage2.jsonl'}, {'split': 'validation', 'path': 'val_train_80.jsonl'}, {'split': 'synthetic_holdout', 'path': 'holdout_synth_400_frozen.jsonl'}, {'split': 'repo-style_holdout', 'path': 'holdout_repo_style_100_frozen.jsonl'}]
Examples of data_files in YAML:
data_files: data.csv
data_files: data/*.png
data_files:
- part0/*
- part1/*
data_files:
- split: train
path: train/*
- split: test
path: test/*
data_files:
- split: train
path:
- train/part1/*
- train/part2/*
- split: test
path: test/*
PS: some symbols like dashes '-' are not allowed in split namesNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
OneDesk β Frozen company-knowledge benchmark
Overview
OneDesk studies a small multilingual assistant that answers current company-knowledge questions across a nine-server MCP environment: notion-mcp, github-mcp, slack-mcp, db-mcp, time-mcp, unit-converter-mcp, math-mcp, weather-mcp, and calendar-mcp.
Structured Python facts determine gold answers before language rendering. Tasks cover consistent settings, stale runbooks, incident cost/configuration chains, and missing-entity or unplugged-source abstention. Scope-aware retrieval mixes production records with newer [STAGING-ONLY] and sibling distractors. SQLite SELECT and deterministic utility calculations execute as real local tools.
The six JSONL files below are stored at the repository root alongside world snapshots and SHA256SUMS.
Companion split table β 1,700 curated rendered tasks
| Companion benchmark split | Rows | Semantic cases |
|---|---|---|
| train_sft_240.jsonl | 240 | 60 |
| train_grpo_stage1.jsonl | 400 | 100 |
| train_grpo_stage2.jsonl | 480 | 120 |
| val_train_80.jsonl | 80 | 20 |
| holdout_synth_400_frozen.jsonl | 400 | 100 |
| holdout_repo_style_100_frozen.jsonl | 100 | 26 |
The train world has 60 services, 80 incidents and 320 semantic cases: 300 allocated plus 20 reserved. Training phases, validation and holdouts have disjoint case IDs, service-parameter pairs and incident IDs. Script styles are en:latin, hi:native, hi:latin, hinglish:latin, te:native and te:latin.
Curation and verification
CuratorKIT 0.1.21 processed 1,955 candidates, retained 1,700, and logged 255 rejected injected exact duplicates. This validates exact deduplication, not semantic paraphrase filtering. The maximum SFT trajectory measured 2,359 tokens; the 2,560-token training budget preserves all 240 oracle conversations without truncation. Training/evaluation never regenerates the frozen inputs.
Schema
| Field | Meaning |
|---|---|
task_id, case_id |
Unique rendered query and shared semantic-case identifier |
world_split, pool, schema_version |
World, phase case pool, schema version 2 |
question, language, script |
User text, en/hi/hinglish/te register and native/Latin script |
prompt |
System instructions, active schemas and user query; model-visible input |
messages |
Full oracle SFT conversation, including tool calls, observations and final response, when present |
kind, incident_subtype |
consistent/conflict/incident/abstain; SLA, payload or config-audit subtype |
service_name, param_name, incident_id |
Entity/parameter or incident selectors |
active_mcps, required_mcps, distractor_mcps |
Available tool interfaces, gold-required chain and routing distractors |
expected_status, gold_value, gold_cites |
Deterministic target status, normalized scalar and supporting IDs |
stale_doc_id, stale_trap_value |
Stale page identifier and outdated scalar trap |
oracle_optimal_steps, reference_tool_sequence |
Measured discovery-policy call length and reference sequence |
Run records additionally include raw assistant turns, messages, calls/outputs, retrieved IDs, predictions, completion reasons, rewards, token masks, group eligibility and optimizer/validation logs. Record types vary by historical run version. The oracle length is a measured policy length, not a proof of global optimality.
Tool protocol:
<tool_call>{"server":"db-mcp","tool":"query_db","arguments":{"sql":"SELECT incident_id FROM production_incidents LIMIT 1"}}</tool_call>
<final_answer>{"status":"answered","value":16,"cites":["retrieved-id"],"stale_source":null}</final_answer>
Checksums and provenance
The frozen data directory has 19 locked files, including a 15-artifact SHA256SUMS. The split hashes are:
| JSONL artifact | SHA-256 |
|---|---|
| train_sft_240.jsonl | 96402f7d8ce8f05ee98aee212f1af9ac7a2b51fef1098e2beda3e395b90e1932 |
| train_grpo_stage1.jsonl | b44e67d78d2abc921583c535a10c3a2bd790aca8fd43723e87151c0c962b7768 |
| train_grpo_stage2.jsonl | 8d7f5236f1fad9450f5ae3ecb81e83b9acda96f01d548225787b7f7eb5754977 |
| val_train_80.jsonl | 4eb608f1391596afe2ad62f67e9c5a68e5b396838b8c97f130638d9510f3b254 |
| holdout_synth_400_frozen.jsonl | 82b92badc6f1fdc71bba41a876db4adf9f19633108b26331297227a7587ac50b |
| holdout_repo_style_100_frozen.jsonl | c6d808482bb85143b08eda51e6703f13f9f0817dd0b30a90ca3cd2d81e8b3469 |
Repository owner: @Revanth-ml. Frozen data Β· Final study artifacts.
Limitations
Facts and repo-style fixtures are synthetic. Configuration values are split by modulo-3 residue; template translations are correlated. Reviewed holdouts and repeated validation selection do not constitute a fresh confirmation study. No real company data, live Notion/Slack/GitHub network connector, or independently verified OSS drift transfer is claimed.
Licensing
Our original synthetic benchmark data and original dataset annotations are licensed under CC-BY-4.0. Attribution: OneDesk contributors / Revanth-ml. Original project code is licensed under Apache-2.0; model weights and adapters retain CC BY-NC 4.0, and bundled third-party code retains its own license. These terms are recorded outside the immutable data files; benchmark contents and checksums are unchanged. CC-BY-4.0 terms: https://creativecommons.org/licenses/by/4.0/. Apache-2.0 terms: https://www.apache.org/licenses/LICENSE-2.0.
Published collections
Model Collection: Revanth-ml/onedesk-models-and-experimental-controls
Data Collection: Revanth-ml/onedesk-data-and-experiment-evidence
- Downloads last month
- -