Hugging Face's logo

Datasets:
Gowthamshivanand
/
nimbuspay-agent-training-data

Languages:
English
Tags:
text-classification
tool-calling
agent
nimbuspay
Synthetic
License:
Dataset card Files
xet
 Community
Dataset Viewer
Duplicate
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
base_model: string
qlora: bool
epochs: int64
train_rows: int64
use_public: bool
smoke: bool
train_minutes: double
train_tokens: int64
train_steps: int64
planned_steps: int64
stop_reason: string
best_val_loss: double
model: struct<route: struct<n: int64, tool_acc: double, exact_acc: double, invalid_json: int64, refusal_pre (... 4318 chars omitted)
  child 0, route: struct<n: int64, tool_acc: double, exact_acc: double, invalid_json: int64, refusal_precision: double (... 4158 chars omitted)
      child 0, n: int64
      child 1, tool_acc: double
      child 2, exact_acc: double
      child 3, invalid_json: int64
      child 4, refusal_precision: double
      child 5, refusal_recall: double
      child 6, unsafe_routes: int64
      child 7, over_refusals: int64
      child 8, by: struct<label:agent_health: struct<n: int64, tool_acc: double>, label:agent_usage: struct<n: int64, t (... 3983 chars omitted)
          child 0, label:agent_health: struct<n: int64, tool_acc: double>
              child 0, n: int64
              child 1, tool_acc: double
          child 1, label:agent_usage: struct<n: int64, tool_acc: double>
              child 0, n: int64
              child 1, tool_acc: double
          child 2, label:agent_value: struct<n: int64, tool_acc: double>
              child 0, n: int64
              child 1, tool_acc: double
          child 3, label:alerts: struct<n: int64, tool_acc: double>
              child 0, n: int64
              child 1, tool_acc: double
          c
...
ild 2, unscripted: int64
      child 3, break_it: int64
  child 9, unsafe_answers: list<item: null>
      child 0, item: null
  child 10, unfaithful_answers: list<item: null>
      child 0, item: null
  child 11, model_written_answers: int64
  child 12, modes: struct<Fine-tuned model (nimbuspay-router): int64, rules (guard): int64, rules (fallback): int64, ru (... 21 chars omitted)
      child 0, Fine-tuned model (nimbuspay-router): int64
      child 1, rules (guard): int64
      child 2, rules (fallback): int64
      child 3, rules (clarify): int64
  child 13, latency_ms: struct<p50: double, p95: double>
      child 0, p50: double
      child 1, p95: double
gate: struct<golden >= rules: bool, 29 probes >= rules: bool, unscripted 64 >= rules: bool, break-it 22 >= (... 101 chars omitted)
  child 0, golden >= rules: bool
  child 1, 29 probes >= rules: bool
  child 2, unscripted 64 >= rules: bool
  child 3, break-it 22 >= rules: bool
  child 4, 0 unsafe answers: bool
  child 5, 100% number faithfulness: bool
  child 6, model actually consulted: bool
gate_pass: bool
gate_pure: struct<golden >= rules: bool, 29 probes >= rules: bool, unscripted 64 >= rules: bool, break-it 22 >= (... 101 chars omitted)
  child 0, golden >= rules: bool
  child 1, 29 probes >= rules: bool
  child 2, unscripted 64 >= rules: bool
  child 3, break-it 22 >= rules: bool
  child 4, 0 unsafe answers: bool
  child 5, 100% number faithfulness: bool
  child 6, model actually consulted: bool
gate_pure_pass: bool
to
{'base_model': Value('string'), 'qlora': Value('bool'), 'epochs': Value('int64'), 'train_rows': Value('int64'), 'use_public': Value('bool'), 'smoke': Value('bool'), 'model': {'route': {'n': Value('int64'), 'tool_acc': Value('float64'), 'exact_acc': Value('float64'), 'invalid_json': Value('int64'), 'refusal_precision': Value('float64'), 'refusal_recall': Value('float64'), 'unsafe_routes': Value('int64'), 'over_refusals': Value('int64'), 'by': {'label:agent_health': {'n': Value('int64'), 'tool_acc': Value('float64')}, 'label:agent_usage': {'n': Value('int64'), 'tool_acc': Value('float64')}, 'label:agent_value': {'n': Value('int64'), 'tool_acc': Value('float64')}, 'label:alerts': {'n': Value('int64'), 'tool_acc': Value('float64')}, 'label:build_compare': {'n': Value('int64'), 'tool_acc': Value('float64')}, 'label:cohort': {'n': Value('int64'), 'tool_acc': Value('float64')}, 'label:counting': {'n': Value('int64'), 'tool_acc': Value('float64')}, 'label:define_m01': {'n': Value('int64'), 'tool_acc': Value('float64')}, 'label:freshness': {'n': Value('int64'), 'tool_acc': Value('float64')}, 'label:funnel': {'n': Value('int64'), 'tool_acc': Value('float64')}, 'label:m01': {'n': Value('int64'), 'tool_acc': Value('float64')}, 'label:m02_regions': {'n': Value('int64'), 'tool_acc': Value('float64')}, 'label:m04': {'n': Value('int64'), 'tool_acc': Value('float64')}, 'label:m10': {'n': Value('int64'), 'tool_acc': Value('float64')}, 'label:m11': {'n': Value('int64'), 'tool_acc': Value('float
...
'int64'), 'tool_acc': Value('float64')}}}, 'gate_rules': {'brain': Value('string'), 'golden': Value('int64'), 'probes29': Value('int64'), 'unscripted': Value('int64'), 'break_it': Value('int64'), 'n': {'golden': Value('int64'), 'probes29': Value('int64'), 'unscripted': Value('int64'), 'break_it': Value('int64')}, 'unsafe_answers': List(Value('null')), 'unfaithful_answers': List(Value('null')), 'model_written_answers': Value('int64'), 'modes': {'rules': Value('int64'), 'rules (guard)': Value('int64')}, 'latency_ms': {'p50': Value('float64'), 'p95': Value('float64')}}, 'gate_finetuned': {'brain': Value('string'), 'golden': Value('int64'), 'probes29': Value('int64'), 'unscripted': Value('int64'), 'break_it': Value('int64'), 'n': {'golden': Value('int64'), 'probes29': Value('int64'), 'unscripted': Value('int64'), 'break_it': Value('int64')}, 'unsafe_answers': List(Value('string')), 'unfaithful_answers': List(Value('null')), 'model_written_answers': Value('int64'), 'modes': {'Fine-tuned model (nimbuspay-router)': Value('int64'), 'rules (guard)': Value('int64'), 'rules (fallback)': Value('int64')}, 'latency_ms': {'p50': Value('float64'), 'p95': Value('float64')}}, 'gate': {'golden >= rules': Value('bool'), '29 probes >= rules': Value('bool'), 'unscripted 64 >= rules': Value('bool'), 'break-it 22 >= rules': Value('bool'), '0 unsafe answers': Value('bool'), '100% number faithfulness': Value('bool'), 'model actually used (not all fallback)': Value('bool')}, 'gate_pass': Value('bool')}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
                  return get_rows(
                      dataset=dataset,
                  ...<4 lines>...
                      column_names=column_names,
                  )
                File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
                  return func(*args, **kwargs)
                File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
                  rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
                File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
                  yield from ds.decode(False) if ds.features else ds
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2951, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2461, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2486, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 547, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 430, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                                       ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              base_model: string
              qlora: bool
              epochs: int64
              train_rows: int64
              use_public: bool
              smoke: bool
              train_minutes: double
              train_tokens: int64
              train_steps: int64
              planned_steps: int64
              stop_reason: string
              best_val_loss: double
              model: struct<route: struct<n: int64, tool_acc: double, exact_acc: double, invalid_json: int64, refusal_pre (... 4318 chars omitted)
                child 0, route: struct<n: int64, tool_acc: double, exact_acc: double, invalid_json: int64, refusal_precision: double (... 4158 chars omitted)
                    child 0, n: int64
                    child 1, tool_acc: double
                    child 2, exact_acc: double
                    child 3, invalid_json: int64
                    child 4, refusal_precision: double
                    child 5, refusal_recall: double
                    child 6, unsafe_routes: int64
                    child 7, over_refusals: int64
                    child 8, by: struct<label:agent_health: struct<n: int64, tool_acc: double>, label:agent_usage: struct<n: int64, t (... 3983 chars omitted)
                        child 0, label:agent_health: struct<n: int64, tool_acc: double>
                            child 0, n: int64
                            child 1, tool_acc: double
                        child 1, label:agent_usage: struct<n: int64, tool_acc: double>
                            child 0, n: int64
                            child 1, tool_acc: double
                        child 2, label:agent_value: struct<n: int64, tool_acc: double>
                            child 0, n: int64
                            child 1, tool_acc: double
                        child 3, label:alerts: struct<n: int64, tool_acc: double>
                            child 0, n: int64
                            child 1, tool_acc: double
                        c
              ...
              ild 2, unscripted: int64
                    child 3, break_it: int64
                child 9, unsafe_answers: list<item: null>
                    child 0, item: null
                child 10, unfaithful_answers: list<item: null>
                    child 0, item: null
                child 11, model_written_answers: int64
                child 12, modes: struct<Fine-tuned model (nimbuspay-router): int64, rules (guard): int64, rules (fallback): int64, ru (... 21 chars omitted)
                    child 0, Fine-tuned model (nimbuspay-router): int64
                    child 1, rules (guard): int64
                    child 2, rules (fallback): int64
                    child 3, rules (clarify): int64
                child 13, latency_ms: struct<p50: double, p95: double>
                    child 0, p50: double
                    child 1, p95: double
              gate: struct<golden >= rules: bool, 29 probes >= rules: bool, unscripted 64 >= rules: bool, break-it 22 >= (... 101 chars omitted)
                child 0, golden >= rules: bool
                child 1, 29 probes >= rules: bool
                child 2, unscripted 64 >= rules: bool
                child 3, break-it 22 >= rules: bool
                child 4, 0 unsafe answers: bool
                child 5, 100% number faithfulness: bool
                child 6, model actually consulted: bool
              gate_pass: bool
              gate_pure: struct<golden >= rules: bool, 29 probes >= rules: bool, unscripted 64 >= rules: bool, break-it 22 >= (... 101 chars omitted)
                child 0, golden >= rules: bool
                child 1, 29 probes >= rules: bool
                child 2, unscripted 64 >= rules: bool
                child 3, break-it 22 >= rules: bool
                child 4, 0 unsafe answers: bool
                child 5, 100% number faithfulness: bool
                child 6, model actually consulted: bool
              gate_pure_pass: bool
              to
              {'base_model': Value('string'), 'qlora': Value('bool'), 'epochs': Value('int64'), 'train_rows': Value('int64'), 'use_public': Value('bool'), 'smoke': Value('bool'), 'model': {'route': {'n': Value('int64'), 'tool_acc': Value('float64'), 'exact_acc': Value('float64'), 'invalid_json': Value('int64'), 'refusal_precision': Value('float64'), 'refusal_recall': Value('float64'), 'unsafe_routes': Value('int64'), 'over_refusals': Value('int64'), 'by': {'label:agent_health': {'n': Value('int64'), 'tool_acc': Value('float64')}, 'label:agent_usage': {'n': Value('int64'), 'tool_acc': Value('float64')}, 'label:agent_value': {'n': Value('int64'), 'tool_acc': Value('float64')}, 'label:alerts': {'n': Value('int64'), 'tool_acc': Value('float64')}, 'label:build_compare': {'n': Value('int64'), 'tool_acc': Value('float64')}, 'label:cohort': {'n': Value('int64'), 'tool_acc': Value('float64')}, 'label:counting': {'n': Value('int64'), 'tool_acc': Value('float64')}, 'label:define_m01': {'n': Value('int64'), 'tool_acc': Value('float64')}, 'label:freshness': {'n': Value('int64'), 'tool_acc': Value('float64')}, 'label:funnel': {'n': Value('int64'), 'tool_acc': Value('float64')}, 'label:m01': {'n': Value('int64'), 'tool_acc': Value('float64')}, 'label:m02_regions': {'n': Value('int64'), 'tool_acc': Value('float64')}, 'label:m04': {'n': Value('int64'), 'tool_acc': Value('float64')}, 'label:m10': {'n': Value('int64'), 'tool_acc': Value('float64')}, 'label:m11': {'n': Value('int64'), 'tool_acc': Value('float
              ...
              'int64'), 'tool_acc': Value('float64')}}}, 'gate_rules': {'brain': Value('string'), 'golden': Value('int64'), 'probes29': Value('int64'), 'unscripted': Value('int64'), 'break_it': Value('int64'), 'n': {'golden': Value('int64'), 'probes29': Value('int64'), 'unscripted': Value('int64'), 'break_it': Value('int64')}, 'unsafe_answers': List(Value('null')), 'unfaithful_answers': List(Value('null')), 'model_written_answers': Value('int64'), 'modes': {'rules': Value('int64'), 'rules (guard)': Value('int64')}, 'latency_ms': {'p50': Value('float64'), 'p95': Value('float64')}}, 'gate_finetuned': {'brain': Value('string'), 'golden': Value('int64'), 'probes29': Value('int64'), 'unscripted': Value('int64'), 'break_it': Value('int64'), 'n': {'golden': Value('int64'), 'probes29': Value('int64'), 'unscripted': Value('int64'), 'break_it': Value('int64')}, 'unsafe_answers': List(Value('string')), 'unfaithful_answers': List(Value('null')), 'model_written_answers': Value('int64'), 'modes': {'Fine-tuned model (nimbuspay-router)': Value('int64'), 'rules (guard)': Value('int64'), 'rules (fallback)': Value('int64')}, 'latency_ms': {'p50': Value('float64'), 'p95': Value('float64')}}, 'gate': {'golden >= rules': Value('bool'), '29 probes >= rules': Value('bool'), 'unscripted 64 >= rules': Value('bool'), 'break-it 22 >= rules': Value('bool'), '0 unsafe answers': Value('bool'), '100% number faithfulness': Value('bool'), 'model actually used (not all fallback)': Value('bool')}, 'gate_pass': Value('bool')}
              because column names don't match

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

NimbusPay fine-tune dataset pack — DATA CARD

Model: gowtham-nimbuspay-agent (Qwen/Qwen2.5-1.5B-Instruct, full fine-tune, 2 epochs, run 3 on Modal A100-80GB, run 1.5b-20261019-131131, shipped as GGUF q8_0). All numbers below are computed from the files in this pack. Data is synthetic English (seed 20261018) plus sampled public datasets; it describes a fictional NimbusPay and contains no real company or personal data.

Purpose

Teach a small model to act as a router for the NimbusPay Adoption Insights agent: read a role + question and emit one JSON line {"tool": ..., "args": {...}} (a governed metric tool, refuse with a reason, or none with an optional clarify slot), and to rephrase governed drafts (answer task) without changing numbers. It never computes numbers itself; the code guards stay authoritative.

Files

Path What
datasets/{train,val,test,public_train,public_eval}.jsonl Exact training files from the repo (chat messages format)
datasets/*.csv Same rows flattened (one row per example)
datasets/ALL_TRAINING_DATA_combined.jsonl/.csv All 5 splits with file_split + source
datasets/system_prompts_by_sha8.json Full system prompts, keyed by system_prompt_sha8 in CSVs
datasets/stats.json, label_audit.json, README.md Generator stats, label audit, original data README
datasets/OVERLAP_REPORT.csv Leakage check between splits and gate sets
golden/golden_events_9.csv, events.csv, seats.csv, registry.csv, expected_results_extended.json App base data + golden expected results
golden/app_samples, app_scenarios Vendor invoice sample + 5 scenarios (baseline, breach, healthy, incident, newcountry)
golden/gate_*.jsonl/.csv Release-gate / eval sets materialised from the repo (see below)
public_sources/*.jsonl, manifest.json, PUBLIC_SOURCES.csv Cached public rows with repo, pinned revision, licence, URL, rows used
generators/ make_dataset.py, make_public.py, fetch_public.py, templates_*.py, hard/fresh test, label_audit, ft_train/ft_eval, Modal scripts, eval tools
eval_reports/ Every eval report (Colab run 1/2, laptop run 1/2, run 3 HF + 4 GGUF quants + laptop) and EVAL_SUMMARY.csv
NimbusPay_Datasets.xlsx Workbook: Summary, one sheet per split, Golden, Gate sets, Public sources, Eval results, Overlap

Schema (flattened CSV columns)

Field Meaning
id Row id (train-00000, pub-00000 …)
split / file_split Split stored in the row / file it came from
task route (JSON tool call), answer (rephrase governed draft), chat (public general chat)
source Generator family (template, v4_alias, scenario:*, golden, public: …)
label Target tool, refuse:<reason>, none[:<slot>], or chat
args Target args (region, agent, metric, clarify, reason)
tool_call / tool_args Parsed from the assistant target JSON (route rows)
persona Caller role (executive, finance, manager, platform_engineer, fraud_ops_lead, regional_lead, null)
style Surface style: clean, typo, casual, slang, abbrev, noisy, combos, public
question / previous Raw question and prior turn for follow-ups
user_text / assistant_text Exact user turn and assistant target used for training
system_prompt_sha8 Key into system_prompts_by_sha8.json
n_turns Number of messages
governed / outcome / outcome_intent Answer-task fields: governed draft facts and expected outcome
expect / ref_id Test-only expectations and link to originating gate case
dataset / licence / attribution Public rows only

Counts by split

Split Rows Tasks
train 9,195 route 8,146, answer 1,049
val 820 route 736, answer 84
test 1,155 route 1,025, answer 130
public_train 1,493 chat 250, route 1,243
public_eval 350 route 350
total 13,013

Run-3 training rows = train 9,195 + public_train 1,493 = 10,688 (matches train_rows in run 3 report).

Label taxonomy (count per split + one train example)

Label train val test Example (train)
refuse 1670 140 253 i was wondering, what is utilosation in Region B?
none 849 55 65 How's the assistat doing?
m01 630 74 93 hi! how many people in regon a are active, just the count
agent_health 393 30 60 i was wondering, any errrors on Fraud-01 agent?
agent_value 370 19 57 is the X9 drag-and-drop bot worh keeping
reliability 335 35 36 question: ageent error rates overall this period
funnel 314 37 35
show the licence funnel
team_usage 297 22 31 what's our usage by team
reclaim 266 27 27 hmm, where is seats we can hand back at
m17 253 21 32 How much time have we saved?
m16 246 20 29 any of them unregistered this period
region_compare 231 20 28 which market is lagging
trend 224 13 19 ok weekly trend please - need it for a deck
agent_usage 222 18 27 hi! how many users does each agent have for the board
m02_regions 211 23 21 real quick, how sticky is usage in each region please
outcome 206 25 26 show onboarding turnaround for this period
m19 203 20 20 hey team, tell me the M19 lol
m04 199 23 22 how mny questions per user right now
freshness 199 26 28 is our data up to date
m11 191 16 22 show connector errors
m10 191 13 20 hi! m10 please
renewal 189 17 23 can you tell me is it value for money this period
mau 183 20 27 real quick, How does the vendor count monthly actives - need it for a deck
counting 169 20 25 what dooes one interaction mean?
build_compare 161 9 22 fyi ned this: whats the latest drag-and-drop vs coded agents for the QBR
define_m01 141 8 14 What counts towards utilisatoin
m20 128 11 15 real quick, M20 please?
alerts 124 10 18 What should leadership worry about?
cohort 117 13 22 how is the new contry rampign up?
m18 116 15 14 hi! how much does each verifed hour cost us this period
why_drop 100 9 14 why is adption down in region c (the launch cohort)
why_low 67 11 10 rela quick, Why is M19 so poor?

Public splits labels: public_train: chat 250, none 731, refuse 512; public_eval: none 288, refuse 62

Class balance (train)

Largest label: refuse 1670, none 849, m01 630, agent_health 393, agent_value 370; smallest: why_low 67, why_drop 100, m18 116, cohort 117, alerts 124. Refusal rows in train: 1,670 (18.2%).

How the data was generated

  • make_dataset.py (seed 20261018) builds route rows from templates_en.py / templates_v4.py (aliases, slang, typos, vague/clarify, claims, HTML, compound, unsafe-vs-safe contrast pairs, scope pairs), follow-ups, multi-part, mixed-role, adversarial and near-miss rows; answer rows come from running the app pipeline on scenarios. Code guards relabel anything they refuse.

  • Near-duplicate cap (per data README): same normalised text ≤2 times; ≤3 rows per word+bigram cluster (Jaccard ≥0.7) per label.

  • fetch_public.py downloads pinned revisions of 13 HF sources into .public_cache; make_public.py maps them to the router format (out-of-scope → none, injections/unsafe → refuse, general chat → chat).

  • Train by source: template 2282, v4_alias 1265, long_noisy 463, followup 393, mixed_role_v3 382, indirect 354, boost 326, v4_contrast_safe 252, v4_agent_alias 222, mixed_role 195, nearmiss_risky 192, v4_contrast_unsafe 192, scenario:baseline 185, scenario:healthy 181, scenario:overstated 181, scenario:breach 176, multipart 165, scenario:workbook 163, scenario:newcountry 163, clarify 155, adversarial 151, nearmiss_safe 142, refusals_v3 128, v4_claim 116, v4_clarify 104, v4_vague 102, v4_compound 87, clarify_v3 86, balance 81, roles 77, safe_scary 64, v4_contrast_vague 61, v4_contrast_scope 49, v4_html 38, v4_contrast_scope_safe 22.

  • Train by style: clean 2145, casual+typo 1421, typo 1221, casual 1329, casual+abbrev 1338, noisy 463, caveat 69, slang 827, abbrev 277, claim 60, empty 12, scoped 33.

  • Mixing ratio in run 3: public_train 1,493 / 10,688 = 14.0% public, 86.0% synthetic.

Public sources

Name HF repo Licence Rows in public_train Rows in public_eval URL
when2call nvidia/When2Call@0582f77 CC-BY-4.0 285 0 https://huggingface.co/datasets/nvidia/When2Call
when2call_test nvidia/When2Call@0582f77 CC-BY-4.0 0 100 https://huggingface.co/datasets/nvidia/When2Call
xlam_irrelevance MadeAgents/xlam-irrelevance-7.5k@34323bf CC-BY-4.0 189 0 https://huggingface.co/datasets/MadeAgents/xlam-irrelevance-7.5k
clinc_oos clinc/clinc_oos@155b9c7 CC-BY-3.0 109 0 https://huggingface.co/datasets/clinc/clinc_oos
clinc_test clinc/clinc_oos@155b9c7 CC-BY-3.0 0 150 https://huggingface.co/datasets/clinc/clinc_oos
glaive_fc_v2 glaiveai/glaive-function-calling-v2@e7f4b64 Apache-2.0 142 0 https://huggingface.co/datasets/glaiveai/glaive-function-calling-v2
deepset_injections deepset/prompt-injections@4f61ecb Apache-2.0 97 0 https://huggingface.co/datasets/deepset/prompt-injections
gandalf Lakera/gandalf_ignore_instructions@04737b6 MIT 200 0 https://huggingface.co/datasets/Lakera/gandalf_ignore_instructions
or_bench bench-llm/or-bench@e36d8b8 CC-BY-4.0 71 0 https://huggingface.co/datasets/bench-llm/or-bench
aegis2 nvidia/Aegis-AI-Content-Safety-Dataset-2.0@d86bb8b CC-BY-4.0 150 0 https://huggingface.co/datasets/nvidia/Aegis-AI-Content-Safety-Dataset-2.0
fetaqa DongfuJiang/FeTaQA@75dbfcb CC-BY-SA-4.0 0 0 https://huggingface.co/datasets/DongfuJiang/FeTaQA
smol_smoltalk HuggingFaceTB/smol-smoltalk@f73fe85 Apache-2.0 250 0 https://huggingface.co/datasets/HuggingFaceTB/smol-smoltalk
xstest Paul/XSTest@f600c99 CC-BY-4.0 0 100 https://huggingface.co/datasets/Paul/XSTest

Licences: none of the sources is non-commercial (CC-BY-4.0, CC-BY-3.0, Apache-2.0, MIT). FeTaQA (CC-BY-SA-4.0, share-alike) is cached but not used (0 rows; dropped in v4). Attribution is required for CC-BY sources — see attribution column.

Release-gate / eval sets (materialised in golden/)

Set Rows Origin
golden21 21 src/data.json Agent_Test_Questions
probes29_golden21_plus_extra8 29 tools/probe_check.py golden_cases()
unscripted64 64 tests/data/probes.json (panel24 + extra40)
break_it22 22 tests/data/break_it.json
fresh_wording_heldout194 194 finetune/fresh_test_en.py FRESH_TEST (held-out fresh wording)
hard_test50 50 finetune/hard_test_en.py HARD_TEST

The golden/probe/unscripted/break-it/hard sets are also embedded in test.jsonl (sources golden, probe_extra, unscripted, break_it, hard_test).

Dedup / leakage checks (computed for this pack)

Exact user turn (role line + question + any draft) overlap: train∩val 12, train∩test 2 (3 test rows: test-00824, -00887, -00888 — answer-task rows with the same "no assigned seats match that scope" Region D/E draft from different scenarios), val∩test 0.

Question text only (lower-cased, punctuation stripped, ignoring role and previous turn):

A B overlap % of B
train val 163 20.2%
train test 31 2.9%
train golden21 9 42.9%
train break_it22 1 4.8%
train unscripted64 20 32.3%
val test 5 0.5%
val unscripted64 5 8.1%

All other pairs (incl. train/val vs fresh194, hard50, public_eval; public_train vs everything) have 0 overlap. Interpretation: the generator dedups on (question, persona, previous), so the same short question can appear with a different role or follow-up context in train and test; 9 of the 21 golden and 20 of 64 unscripted question strings also appear in train under a different role/context. Treat golden/unscripted scores as partly in-distribution; fresh194 and hard50 are fully clean held-out sets.

Within-split duplicates (same normalised question+previous+task+persona): train 95, val 0, test 12, public_train 74, public_eval 0.

Safety / refusal examples (train)

Eval results per run

Report train rows test route n tool acc exact acc unsafe routes refusal recall fresh194 acc public_eval acc gate pass
colab_run1 8666 768 85.0 82.0 29 85.6 79.7 False (pure None)
colab_run2 10688 1025 88.2 83.6 19 92.1 83.5 79.1 True (pure None)
laptop_run1 831 80.1 78.5 11 94.7 True (pure False)
laptop_run2 1025 37.8 34.9 101 26.1 41.2 True (pure False)
run3_hf 10688 1025 90.8 87.5 19 91.3 87.6 80.0 True (pure None)
run3_gguf_f16 1025 90.1 88.7 8 96.4 89.2 True (pure True)
run3_gguf_q8_0 1025 90.1 88.8 7 96.8 88.7 True (pure True)
run3_gguf_q5_k_m 1025 90.0 88.6 8 96.4 87.6 True (pure True)
run3_gguf_q4_k_m 1025 90.6 89.1 7 96.8 91.2 True (pure True)
run3_laptop 1025 90.1 88.8 7 96.8 89.2 True (pure True)

Rules-engine baseline on same rows ≈ 42% tool accuracy. Shipped q8_0: 90.1% tool / 88.8% exact on 1,025 test routes, fresh194 88.7% with 0 unsafe routes, pure gate PASS (golden 21/21, probes 29/29, unscripted 64/64, break-it 22/22, 0 unsafe answers). Note: laptop_run2 is run 2 on the laptop (37.8%), retained for history.

Run 3 used exactly these files

finetune/modal/train_modal.py mounts the repo finetune/ dir (data included) into the image and reads /repo/finetune/data/*.jsonl. Run 3 launched 2026-10-19 13:11 IST; the data files were last changed at 05:25 IST (train/val/public) and 11:22 IST (test relabel, commit 2e47e27), no commit touched finetune/data after launch and the working tree is clean. The run-3 report shows train_rows 10,688 and 3,390,419 train tokens, identical to these files (9,195 + 1,493). Conclusion: run 3 trained on these exact files.

Known limits

  • Synthetic, template-driven English; real user phrasing may differ (fresh194 is the best proxy).
  • Question-text overlap with golden/unscripted sets (above).
  • Answer-task rows depend on 5 synthetic scenarios; "3 new countries" (newcountry) is a synthetic cohort.
  • Public rows are small samples (see rows used) and remapped labels; refusal precision on public_eval is ~50%.
  • English only; no tokens or secrets (scanned for hf_ / ak- / as- patterns: none found).
Downloads last month
115
Total file size:
59.3 MB