mondk/Jev-Decisions-Sft

Source: LocalLLaMA/typed-decisions

The dataset covers four independent workflows, 300 records each (1,200 total). Every record asks for five structured fields about the case — a mix of categorical choices, ordinal scores, and boolean yes/no questions.

• Records: 1,200
• Workflows: 4 × 300
• Target fields: 5 per record → 6,000 labelled fields total
• Language: English
• Format: JSONL (one JSON object per line)

---

Format

{"workflow": "...", "state": "...", "questions": "...", "gold": "...", "factors": "...", "label_agreement": "..."}

state, questions, gold, factors and label_agreement are string-encoded JSON. They are stored as strings (escaped, with \" inside) rather than as nested objects, so a naive read gives you a string, not a dict. You must parse them yourself:

import json

with open("train-00000-of-00001.jsonl", encoding="utf-8") as f:
    for line in f:
        row = json.loads(line)              # line -> dict of 6 string values
        state     = json.loads(row["state"])       # -> nested dict
        questions = json.loads(row["questions"])   # -> nested dict
        gold      = json.loads(row["gold"])        # -> nested dict
        factors   = json.loads(row["factors"])     # -> nested dict
        agreement = json.loads(row["label_agreement"])
        workflow  = row["workflow"]                # already a plain string

With datasets:

from datasets import load_dataset

ds = load_dataset("json", data_files="train-00000-of-00001.jsonl", split="train")
# ds[0]["state"] is a str -> json.loads(...) it

ds = ds.map(lambda r: {
    "state":     json.loads(r["state"]),
    "questions": json.loads(r["questions"]),
    "gold":      json.loads(r["gold"]),
    "factors":   json.loads(r["factors"]),
    "label_agreement": json.loads(r["label_agreement"]),
})

---

Columns

Every line has exactly these six keys.

Column	Type	Size (chars: min / median / max)	What it is	What it is for
workflow	plain string	1 of 4 values	Which of the four domains this record belongs to.	Selecting the right question schema, filtering, per-domain splits and metrics.
state	JSON string → object	226 / 725 / 2,002	The case itself: the material to be judged. Structure is workflow-specific.	This is the model input. Everything the judge is allowed to see.
questions	JSON string → object	1,429 / 1,503 / 1,903	The question schema: for each of the 5 target fields, the instructions, the type, and the criteria that define each option. Identical for all records within a workflow (4 distinct values in the whole file).	This is the task specification. It defines the output contract — field names, allowed values, and the meaning of each option. Put it in the prompt.
gold	JSON string → object	717 / 811 / 924	The gold answers. One entry per target field, containing type, label, probabilities, confidence, and (for score items) score or (for boolean items) noul.	This is the training target. See The gold object below.
factors	JSON string → object	251 / 285 / 394	A pre-digested feature summary of the case, flattened per workflow (e.g. tone, topic, issue, violated_constraint).	Derived annotation layer. Useful for analysis, not recommended as model input — see Leakage.
label_agreement	JSON string → object	439 / 469 / 489	Per-field annotator-agreement statistics: argmax_agree, argmax_majority, total_variation.	Data-quality auditing. Must not be used as model input — see Leakage.

state keys by workflow

workflow	Top-level keys
agent_trace_observability	agent, constraints, task, trace_summary
customer_service	account, thread, orders (present in 167/300)
invoice_processing	invoice, purchase_order, delivery, payment, vendor_history
security_incidents	alert, context, history, principal

factors keys by workflow

workflow	Keys
agent_trace_observability	autonomy, completion, constraints, irreversible_action, steps, task, tool_errors, violated_constraint
customer_service	tier, tenure_months, prior_tickets_90d, turns, tone, topic, amount_usd, money_involved, has_deadline, threatens_to_leave, prior_resolution, text_generated
invoice_processing	invoice_total_usd, po_total_usd, variance_usd, variance_pct, issue, issue_description, delivered, overdue, days_until_due, payment_terms, early_payment_discount, vendor_disputes_12m
security_incidents	rule, principal_type, privileged, mfa_enrolled, rotation_days_ago, prior_alerts_90d, source_is_known, change_window_active, has_benign_explanation, text_generated

---

The gold object

gold maps each of the five target field names to an object:

{
  "action": {
    "type": "choice",
    "label": "human_review",
    "confidence": 0.244444,
    "probabilities": {
      "continue": 0.283333, "human_review": 0.433333,
      "observe": 0.25, "stop": 0.033333
    }
  },
  "risk": {
    "type": "score",
    "label": "0",
    "score": 0.666667,
    "confidence": 0.333333,
    "probabilities": { "0": 0.466667, "1": 0.416667, "2": 0.1, "3": 0.016667 }
  },
  "needs_review": {
    "type": "noul",
    "label": "true",
    "noul": 0.55,
    "confidence": 0.1,
    "probabilities": { "false": 0.45, "true": 0.55 }
  }
}

Three field types occur. Each has a different key set:

type	Count	What label is	Extra key(s) present
choice	1,800	A nominal category, e.g. human_review, approve, investigate	—
score	2,400	An ordinal level as a string: "0", "1", "2", … (0 = lowest)	score = expected value of the ordinal level
noul	1,800	A boolean as a string: "true" / "false"	noul = P(true)

Semantics of the derived keys

These are all deterministic functions of probabilities — they carry no information that probabilities does not already have. Verified against all 6,000 fields:

Key	Definition	Verified
probabilities	The normalised distribution over the options: non-negative values summing to 1 (up to ~1e-6 rounding), stored with up to 6 decimal places.	—
label	The argmax of probabilities, as a string.	5,946 / 6,000 exact
score	Expected value over the ordinal levels: Σ level × P(level).	2,400 / 2,400
noul	P("true").	1,800 / 1,800
confidence	A 0–1 margin measuring how separated the top option is from the rest.	see below

confidence in detail. It is a derived margin, not an independent annotation:

• for noul (2 options): confidence = p_max − 0.5 — exact on all 1,800 fields
• for choice: confidence = (p_max − 1/k) / (1 − 1/k), where k is the number of options — exact on all 1,800 fields
• for score: a different margin measure over the ordinal distribution; it does not reduce to the formulas above. Treat it as an opaque 0–1 agreement margin.

Why label differs from argmax in 54 fields. In 54 of the 6,000 fields (0.9 %) the top probability is an exact tie, and the generator broke the tie by an internal rule that is not consistent across cases. If you recompute argmax yourself you will disagree with label on those 54 fields. Always use the stored label; do not recompute it.

---

Question schemas

questions is constant within a workflow, so there are only four distinct schemas. Each target field carries instructions (the question), type, and — for choice — a criteria map giving the meaning of every allowed option. Score items carry criteria as an ordered list, index 0 being the lowest level.

agent_trace_observability — 300 records

Field	Type	Options (label distribution)
action	choice	continue 40 · human_review 135 · observe 122 · stop 3
outcome	choice	failure 15 · harmful 62 · partial 141 · success 82
needs_review	noul	false 154 · true 146
risk	score	0–3 (4 levels): 67 / 106 / 65 / 62
urgency	score	0–3 (4 levels): 70 / 152 / 26 / 52

customer_service — 300 records

Field	Type	Options (label distribution)
action	choice	answer_directly 65 · escalate_to_human 167 · execute_refund 9 · request_information 59
category	choice	account 77 · billing 68 · delivery 60 · refund 27 · technical 68
churn_risk	score	0–3 (4 levels): 15 / 74 / 146 / 65
needs_human	noul	false 99 · true 201
urgency	score	0–3 (4 levels): 3 / 61 / 111 / 125

invoice_processing — 300 records

Field	Type	Options (label distribution)
disposition	choice	approve 120 · hold 56 · manual_review 111 · reject 13
matches_order	noul	false 151 · true 149
duplicate	noul	false 272 · true 28
discrepancy_severity	score	0–3 (4 levels): 133 / 3 / 80 / 84
urgency	score	1–3 (3 levels present): 131 / 126 / 43

security_incidents — 300 records

Field	Type	Options (label distribution)
disposition	choice	close_benign 30 · contain 35 · investigate 214 · monitor 21
credential_compromise	noul	false 132 · true 168
true_positive	noul	false 95 · true 205
severity	score	0–4 (5 levels): 6 / 37 / 91 / 133 / 33
urgency	score	0–3 (4 levels): 8 / 53 / 81 / 158

Note that some score scales are shorter than the criteria list implies (e.g. invoice_processing / urgency never takes the value 0 in this split). Derive the option set from criteria, not from the observed labels.

---

Which columns to use for training

Required

Column	Role	How to use it
state	input	Serialise into the prompt. This is the only case-specific information the model may see.
questions	task spec	Serialise into the system prompt (or the user prompt). Defines the field names, the allowed values and the option meanings.
gold	target	The label. For hard targets use gold[field]["label"].

Recommended

Column	Role	How to use it
workflow	metadata	Select the schema, build per-domain splits and report per-domain metrics. Harmless in the prompt.
gold[field]["probabilities"] / ["score"]	soft target	Only if you are training on distributions (soft-label distillation, ordinal regression, calibration). Ignore it for plain SFT.

Do not feed as input

Column	Reason
label_agreement	Direct label leakage. label_agreement[f]["argmax_majority"] equals gold[f]["label"] in 94.9 % of all 6,000 fields (5,694 / 6,000). For several fields it is worse: matches_order 100 %, category 99.7 %, needs_human and duplicate 98.3 %. A model given this column can copy the answer without reading state.
factors	Partial leakage + redundancy. It contains pre-computed conclusions about the case — e.g. issue: "price_variance" (tells you matches_order), has_benign_explanation and violated_constraint (tell you the disposition), threatens_to_leave (tells you churn_risk). It is also a strict subset of state plus these inferences.

If you deliberately want the model to learn to reproduce these columns as targets as well, move them to the output side — never to the input side.

Minimal training example

import json

def to_example(row):
    questions = json.loads(row["questions"])
    state     = json.loads(row["state"])
    gold      = json.loads(row["gold"])

    prompt = (
        "You are given a case and a set of questions. "
        "Answer every question. Reply with a single JSON object mapping each "
        "question id to its chosen label.\n\n"
        f"Questions:\n{json.dumps(questions, indent=2)}\n\n"
        f"Case:\n{json.dumps(state, indent=2)}"
    )
    answer = json.dumps({f: v["label"] for f, v in gold.items()})
    return prompt, answer

For a chat-format dataset, map it as system = json.dumps(questions), user = json.dumps(state), assistant = json.dumps({f: v["label"] for f, v in gold.items()}).

---

ty