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
Files changed (3) hide show
  1. README.md +2 -2
  2. reorgbench/gen/build.py +2 -2
  3. 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 and expert tasks get 1.5× the rounds and 2× the actions of the fastest careful oracle; standard tasks get 3× both. Tight rounds force parallel work; 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.
 
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 on the hard and expert tiers, as a multiple of the oracle's
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((ROUND_BUDGET if task.tier != "standard" else 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))
 
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 and expert tasks, 3x for standard."""
57
  for t in list(GEN.values())[::7]:
58
  res = run_oracle(t)
59
  assert res.verdict.success
60
- mult = 3.0 if t.tier == "standard" else 1.5
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