HassamGani commited on
Commit ·
e8d32cd
1
Parent(s): ed863e3
Expert tier: round cap 1.2x the oracle's, so bulk work has to be split across hires
Browse files- README.md +2 -2
- reorgbench/gen/build.py +2 -2
- tests/test_generator.py +2 -2
README.md
CHANGED
|
@@ -104,7 +104,7 @@ ReOrgBench is one environment and one task generator, used both as a benchmark a
|
|
| 104 |
|
| 105 |
- **Templates, not just seeds.** The test split holds out whole scenario templates, and some task shapes and change types. Splitting by template is what stops a model from scoring by memorizing templates. A small lite split, a stratified subset of test, is meant for cheap runs.
|
| 106 |
- **Filtering.** `uv run reorgbench list --split dev` lists one split. Add `--tier hard` to list or run only the hard tier, `--template` for one scenario template, and `--limit` to cap the count.
|
| 107 |
-
- **Tiers.** Each split has a standard and a hard tier, with the same verifier rules; generated tasks add an expert tier (below). Hard tasks have three phases with compound changes, at least three roles across the episode, and budgets of 1.5× the oracle's rounds and actions (2.7–4× for standard tasks). A hard task's oracle makes the lookups a careful team can't skip, such as reading each ticket to see whose it is, so the budget isn't set by a plan that already knows the answers. They also plant traps that only the data resolves:
|
| 108 |
- near-duplicate records;
|
| 109 |
- a lapsed contract term or an expired quote;
|
| 110 |
- a closed account;
|
|
@@ -162,7 +162,7 @@ From the units, the generator derives per phase:
|
|
| 162 |
|
| 163 |
**References by construction.** References come from the plan, never from a model run.
|
| 164 |
|
| 165 |
-
- **Budgets.** Hard
|
| 166 |
- **Staffing.** Every crew size the oracle can finish within those budgets is an acceptable staffing.
|
| 167 |
- **Triggers.** The first change fires on a record a third of the way through the work. Each later one fires on progress partway through the work the one before set off. It is chosen from a run with it held back, once the team has settled. Fallback rounds come from the same runs.
|
| 168 |
- **Gate.** The task is kept only if the oracle passes every check, every change fires on its trigger, and the oracle hits no tool errors.
|
|
|
|
| 104 |
|
| 105 |
- **Templates, not just seeds.** The test split holds out whole scenario templates, and some task shapes and change types. Splitting by template is what stops a model from scoring by memorizing templates. A small lite split, a stratified subset of test, is meant for cheap runs.
|
| 106 |
- **Filtering.** `uv run reorgbench list --split dev` lists one split. Add `--tier hard` to list or run only the hard tier, `--template` for one scenario template, and `--limit` to cap the count.
|
| 107 |
+
- **Tiers.** Each split has a standard and a hard tier, with the same verifier rules; generated tasks add an expert tier (below). Hard tasks have three phases with compound changes, at least three roles across the episode, and budgets of 1.5× the oracle's rounds and 2× its actions (1.2× rounds on the expert tier, 2.7–4× for standard tasks). A hard task's oracle makes the lookups a careful team can't skip, such as reading each ticket to see whose it is, so the budget isn't set by a plan that already knows the answers. They also plant traps that only the data resolves:
|
| 108 |
- near-duplicate records;
|
| 109 |
- a lapsed contract term or an expired quote;
|
| 110 |
- a closed account;
|
|
|
|
| 162 |
|
| 163 |
**References by construction.** References come from the plan, never from a model run.
|
| 164 |
|
| 165 |
+
- **Budgets.** Hard tasks get 1.5× the rounds of the fastest careful oracle and expert tasks 1.2×; both get 2× its actions, and standard tasks get 3× both. Tight rounds force parallel work: on the expert tier a team that leaves bulk record edits to one agent runs out of rounds; the extra actions leave room for reading. "Careful" means each new hire pages through what its role works from, reads the docs its part depends on (the handbook included), and support reads each ticket before answering it. The token cap is 30,000 tokens per allowed action (at least 2M), because context per call grows with the number of customers.
|
| 166 |
- **Staffing.** Every crew size the oracle can finish within those budgets is an acceptable staffing.
|
| 167 |
- **Triggers.** The first change fires on a record a third of the way through the work. Each later one fires on progress partway through the work the one before set off. It is chosen from a run with it held back, once the team has settled. Fallback rounds come from the same runs.
|
| 168 |
- **Gate.** The task is kept only if the oracle passes every check, every change fires on its trigger, and the oracle hits no tool errors.
|
reorgbench/gen/build.py
CHANGED
|
@@ -34,7 +34,7 @@ HIRE_ORDER = ["support", "finance", "engineer", "procurement", "legal", "analyst
|
|
| 34 |
LOOSE = Budgets(round_cap=400, action_cap=4000)
|
| 35 |
LIST_ROWS = "db.query"
|
| 36 |
HARD_BUDGET, STANDARD_BUDGET = 2.0, 3.0
|
| 37 |
-
ROUND_BUDGET = 1.5 # rounds
|
| 38 |
TOKENS_PER_ACTION = 30_000
|
| 39 |
OWED = "never:owed" # tags a plan step that never runs; it only says who still needs a fact
|
| 40 |
|
|
@@ -1136,7 +1136,7 @@ def calibrate(task: Task, knobs: Knobs) -> tuple[str | None, Any]:
|
|
| 1136 |
R, A = res.rounds, res.actions_used
|
| 1137 |
# rounds stay tight on the harder tiers so the work has to run in parallel; actions get room for careful reading
|
| 1138 |
mult = knobs.budget if task.tier != "standard" else max(knobs.budget, STANDARD_BUDGET)
|
| 1139 |
-
cap = math.ceil(
|
| 1140 |
actions = math.ceil(mult * A)
|
| 1141 |
# big tasks need more context than the default token cap allows; scale it with the actions allowed
|
| 1142 |
task.budgets = Budgets(round_cap=cap, action_cap=actions, token_cap=max(2_000_000, TOKENS_PER_ACTION * actions))
|
|
|
|
| 34 |
LOOSE = Budgets(round_cap=400, action_cap=4000)
|
| 35 |
LIST_ROWS = "db.query"
|
| 36 |
HARD_BUDGET, STANDARD_BUDGET = 2.0, 3.0
|
| 37 |
+
ROUND_BUDGET = {"hard": 1.5, "expert": 1.2} # rounds as a multiple of the oracle's; expert work has to be split
|
| 38 |
TOKENS_PER_ACTION = 30_000
|
| 39 |
OWED = "never:owed" # tags a plan step that never runs; it only says who still needs a fact
|
| 40 |
|
|
|
|
| 1136 |
R, A = res.rounds, res.actions_used
|
| 1137 |
# rounds stay tight on the harder tiers so the work has to run in parallel; actions get room for careful reading
|
| 1138 |
mult = knobs.budget if task.tier != "standard" else max(knobs.budget, STANDARD_BUDGET)
|
| 1139 |
+
cap = math.ceil(ROUND_BUDGET.get(task.tier, mult) * R)
|
| 1140 |
actions = math.ceil(mult * A)
|
| 1141 |
# big tasks need more context than the default token cap allows; scale it with the actions allowed
|
| 1142 |
task.budgets = Budgets(round_cap=cap, action_cap=actions, token_cap=max(2_000_000, TOKENS_PER_ACTION * actions))
|
tests/test_generator.py
CHANGED
|
@@ -53,11 +53,11 @@ def test_every_change_kind_each_template_offers_shows_up():
|
|
| 53 |
|
| 54 |
|
| 55 |
def test_budgets_follow_the_tier():
|
| 56 |
-
"""Round caps are a fixed multiple of the fastest careful plan: 1.5x for hard
|
| 57 |
for t in list(GEN.values())[::7]:
|
| 58 |
res = run_oracle(t)
|
| 59 |
assert res.verdict.success
|
| 60 |
-
mult = 3.0
|
| 61 |
assert t.budgets.round_cap == pytest.approx(mult * res.rounds, abs=1), t.id
|
| 62 |
|
| 63 |
|
|
|
|
| 53 |
|
| 54 |
|
| 55 |
def test_budgets_follow_the_tier():
|
| 56 |
+
"""Round caps are a fixed multiple of the fastest careful plan: 1.5x for hard, 1.2x for expert, 3x for standard."""
|
| 57 |
for t in list(GEN.values())[::7]:
|
| 58 |
res = run_oracle(t)
|
| 59 |
assert res.verdict.success
|
| 60 |
+
mult = {"standard": 3.0, "hard": 1.5, "expert": 1.2}[t.tier]
|
| 61 |
assert t.budgets.round_cap == pytest.approx(mult * res.rounds, abs=1), t.id
|
| 62 |
|
| 63 |
|