Jainamshahhh/hr-ops-tools

HR-Ops: 8,621 rows of tool calling and cited policy for HR assistants

A training set for HR-operations assistants, built around one idea: make the HR task objectively checkable. The headline shard is tool calling against authored HR-ops function schemas, where a correct answer is exact JSON and a wrong one cannot hide behind fluent prose. Built for the Adaption AutoScientist Challenge, Part 2 (HR).

What this dataset proves, and how you check it

rows	8,621, of which 5,000 are tool-calling rows
rows that teach the model not to act	21%: ask for the missing parameter, or decline when no tool applies
labels authored or judged by a language model	zero, targets are computed by the generator
what it does to a model	gemma-3-4b-it: 49.3% to 100.0% exact match on held-out probes
the failure it fixes	the base fabricated a withheld parameter in 31 of 33 cases and reported success
significance	p = 2.6e-23 held out, p = 3.0e-6 on the harder set
what this card volunteers	a deliberately harder probe set where the score is 41.3%, released in full
dataset self-checks	13 of 13 pass, re-derivable with the released verifier
what ships beside the rows	both scorers, the verifier, both probe sets, raw generations, per-row verdicts

Every number above is recomputable on a laptop with no GPU, from files in this repository.

TL;DR

• 8,621 rows built around one idea a lot of HR data misses: make the answer mechanically checkable. The headline shard is 5,000 HR-ops tool calls where a correct answer is exact JSON, so no wrong answer can hide behind fluent prose and no judge has to be trusted.
• The failure cases are 21% of the corpus, not an afterthought. clarify and refuse are first-class training targets. A model trained only on successful calls has no way to learn that declining is an option, which is exactly how a strong base model ends up fabricating a start date in 31 of 33 cases and clarifying in none of them.
• Two held-out probe sets from a disjoint seed space, and we ship the one we score worse on. A 150-row set, plus a deliberately harder 75-row set built because the first shares templates with training. Releasing only the flattering probes would make any headline unfalsifiable.
• 13 of 13 integrity checks pass, re-derivable in seconds on a laptop with no GPU and no API key. The verifier exits non-zero on any failure. Please run it.
• Three answer types, one typed envelope: call, clarify, refuse. A model that always emits a call is forced to fabricate arguments, so the failure cases are first-class training targets rather than an afterthought.
• 961 federal-policy rows that are both accurate and complete: drawn from a pool fact-checked at 98.8%, 100% carrying a source URL, and filtered to the fuller half.
• 115 authored corrections for every fabrication a base model was measured making (FMLA week counts, COBRA deadlines, FLSA tip credit and travel pay, youth wage).
• Zero real personal data. Resume and profile rows are fully synthetic.
• Public-domain and openly-licensed sources only; per-row license and source preserved, with policy rows carrying the exact fact-sheet URL they came from.

Why this shape, stated plainly

An earlier version of this entry reached parity with its base model, and a second attempt regressed. The post-mortem was specific and is worth repeating, because it drove every decision here:

1. Generic HR advisory is not a capability gap. Modern 4B models already write decent job descriptions and interview questions. Training on more of that buys style, not skill, and style needs a judge, and a judge cannot reliably tell a real citation from an invented one. So the headline moved to tool calling, where correctness is mechanical.
2. Accuracy alone made the model worse. The second attempt used more accurate but extractive, short policy answers, and the model generalised that terseness across every task, answering "give me six interview questions" with four. Accuracy went up and usefulness went down. Here, policy rows are selected for completeness inside the accurate pool, rather than trading one against the other.

Composition

Shard	Rows	What it teaches	Source / license
hr_tool_ops	5,000	call / clarify / refuse over 12 HR-ops tools	authored, CC-BY-4.0
hr_policy_qa_cited	961	correct federal rule with a source URL	DOL, EEOC, USCIS fact sheets, public domain
hr_jd_writing	700	complete job descriptions	O*NET 30.3, CC-BY-4.0
hr_interview_gen	700	competency question sets	O*NET 30.3, CC-BY-4.0
hr_resume_fit	700	job-relevant screening, fairness-aware	synthetic from O*NET, CC-BY-4.0
hr_multiwoz_dialog	400	multi-turn HR service dialogue	HR-MultiWOZ, Apache-2.0
hr_corrections	115	explicit corrections of measured fabrications	authored, CC0
hr_refusal_boundary	45	"federal law does not set this"	authored, public domain

Tool-shard answer mix: 3,187 call / 1,034 clarify / 779 refuse. Probe set: 98 / 33 / 19.

Two probe sets, and why the second one exists

The 150-row probe set has zero full-prompt overlap with training and comes from a disjoint seed space, which is the usual bar. It is not a sufficient bar here, and the precise reason is worth stating rather than leaving for a reviewer to find.

A full prompt is a tool menu plus a request line. The menu is a shuffled random subset, so full prompts are all distinct and the zero-overlap assertion is true as written. Strip the menu and compare only the request line, and the picture changes: 44 of the 150 request lines appear verbatim somewhere in the training shard, and similarity to the nearest training request line averages 0.95 (median 0.95, p10 0.88). Both sets come from the same templates, so a high score on this set shows the behaviour was learned, not how far it transfers.

So a second set of 75 rows (36 call / 29 clarify / 10 refuse) is released alongside it, built to break the resemblance: unseen request phrasings, refusal categories that appear nowhere in training, and noisy slot formats. 0 of its 75 request lines appear verbatim in training, and nearest-neighbour similarity drops to a mean of 0.51 (median 0.50, p90 0.61).

Both are shipped. The harder set is the one a model scores worse on, which is precisely why withholding it would be dishonest: releasing only the flattering probes would make the headline number unfalsifiable. Reproduce this comparison yourself with part2/scripts/hr/probe_similarity.py.

The tool-calling contract

Twelve authored HR-ops functions, each with typed and required-marked parameters: request_leave, get_pto_balance, schedule_interview, lookup_policy, file_expense, update_employee_record, start_onboarding, submit_timesheet, search_open_roles, enroll_benefits, book_training, get_org_chart.

Each row shows a random subset of tools (always including the right one, order shuffled) so position cannot be memorised, and asks for exactly one JSON object:

{"action": "call",    "tool": "request_leave", "arguments": {"employee_id": "EMP-1042", "leave_type": "sick", "start_date": "2026-03-02", "end_date": "2026-03-06"}}
{"action": "clarify", "missing": ["start_date"], "question": "Which start date should I use?"}
{"action": "refuse",  "reason": "No available tool supports this request."}

Scoring is exact: JSON validity, action, tool name, and arguments compared by value with key order ignored. A clarify row fails if the model invents arguments while claiming it needs information, which is precisely the fabrication behaviour the earlier attempt suffered from. There is no partial credit for a confident wrong call.

Excluded on purpose: a general function-calling corpus (ToolACE) was available and dropped. Its functions include things like countiesSearchInBoundingBox and anthropology.get_cultural_artifacts, which teach nothing about HR and would dilute the authored schemas. Only ~342 of its rows survived HR keyword filtering, and none were HR-specific in substance.

Generation quality control

Rows are programmatic, so defects are systematic rather than random, and spot-reading found three that a reviewer would also have found. All are now zero, verified by assertion:

• Repeated slots. A slot named in the request stem was appended again in the detail list, producing "Pull up the equipment policy for topic equipment."
• Impossible date ranges. 326 of 5,000 rows had a period ending before it began. Start and end are now ordered, and only when both halves are present, so clarify rows that deliberately withhold one are untouched.
• Redundant units. 744 rows read "duration minutes 45 minutes" or "hours 20.1 hours".

Plus: 0 malformed JSON targets, 0 duplicate prompts, 0 em dashes, and train/probe prompt overlap asserted 0 before writing.

Verify these claims yourself, without a GPU

python part2/scripts/hr/verify_hr_dataset.py

Re-derives every claim on this card from the released files: that each tool row is one JSON object whose action matches its label, that no clarify row carries arguments, that every named tool exists and every argument key is a real parameter of it, that every call supplies all required parameters, that neither probe set shares a prompt with training or with the other, that no hard-probe request line recurs verbatim in training, that no date range is inverted, and that every policy row carries a source URL. Exits non-zero on any failure.

Latest run, all 13 checks pass:

Check	Result
tool rows that are one JSON object with the right action	5,225 / 5,225
clarify rows that invent arguments	0
unknown tool names	0
argument keys outside the tool's parameters	0
calls missing a required parameter	0
prompts shared between training and either probe set	0
prompts shared between the two probe sets	0
hard-probe request lines recurring verbatim in training	0 of 75
inverted start/end date pairs	0
doubled unit phrasings	0
policy rows carrying a source URL	961 / 961
em or en dashes	0

The last row of that table is the one that backs the "two probe sets" section above. The easy set's resemblance to training is a stated limitation; the hard set's independence is an asserted invariant, so it fails the build rather than relying on anyone's word.

Try it, and everything that backs it

Live side by side demo: https://huggingface.co/spaces/Jainamshahhh/opshr-demo Enter your own input and watch the base model and OpsHR-4B answer it under identical greedy decoding. The GPU backend scales to zero, so a cold first request takes about a minute.

	Hugging Face	Kaggle
model	https://huggingface.co/Jainamshahhh/opshr-4b	https://www.kaggle.com/datasets/jainamshahhh/opshr-4b
dataset	https://huggingface.co/datasets/Jainamshahhh/hr-ops-tools	https://www.kaggle.com/datasets/jainamshahhh/hr-ops-tools

Released on both platforms, with the scorer, every eval slice, the per row verdicts and significance.py alongside, so every number on this page can be recomputed rather than trusted.

Intended use and limitations

Fine-tuning small assistants for HR operations and grounded HR policy answers.

• US federal law only. No state, provincial or non-US employment law. The refusal rows exist because the correct answer to "how many sick days must I give" under federal law is that federal law does not set one.
• Not legal advice. Policy rows quote and cite public fact sheets; they are a starting point for an HR professional, not a substitute for counsel. Fact sheets are also amended, so figures should be re-verified against the cited URL.
• The tool schemas are authored, not a real API. They are realistic in shape, but a model trained here will not know your HRIS. Re-target the schemas to your own tools.
• Resume rows are synthetic, which is deliberate for privacy but means they lack the messiness of real applications.
• English only.

Author: Jainam Shah. Attribution: ONET 30.3 (CC-BY-4.0, US Department of Labor), US DOL / EEOC / USCIS fact sheets (public domain), HR-MultiWOZ (Apache-2.0). Tool schemas, corrections and refusal rows authored for this release.*