cmpatino's picture
download
raw
3.74 kB
# ProgramBench, "Team PIV, async" (arXiv 2610.17471, Sec 4): the async team of
# examples/programbench_team_async.yaml (swarmlab repo) with 4 workers, every agent the same
# timeout_s 900: Flint answers with up to 32k reasoning tokens at ~120 tok/s on a shared server; the
# provider's 90 s default timed out 4 solo runs in the 2026-10-22 sweep.
# model, the lead given ../prompts/piv-lead.txt and the workers ../prompts/piv-worker.txt as
# `guidance` (swarmlab's wording of the paper's Plan -> Implement -> Verify regime): the lead
# delegates in all three phases and collects every report (`wait_for_workers`, so the lead needs
# `wait_tool: true`) before moving on. Its pair is solo-piv.yaml. The guidance paths are absolute
# and read at run time; the launcher (another branch) stages such files into the job and
# rewrites the paths.
#
# Step limits: lead 1000 (the official limit), each worker 300. `rounds_per_agent` caps every
# agent's turns, idle ones included: 120. The arithmetic: at 10 steps per turn the lead's 1000
# steps take 100 full turns; 20 more cover turns cut short (an `end_turn` call, the call cap). A
# worker's 300 steps take 30 full turns plus at most one short turn per assignment (its report
# ends the turn); its other turns are idle ones, which make no model call and end only by the
# deadlock rule (every other agent waiting or finished). So no agent's work is cut by the cap,
# and a worker that idles (never assigned, or after the lead stopped) closes at most 120 ticks.
# The run's tick cap is 120 x 5 agents = 600 (`max_rounds`).
#
# The model is a placeholder for a self-hosted server: `served:flint` on the `served` provider.
# `base_url` is a placeholder (an openai_compat provider needs one today); `swarmlab job run
# --endpoint` (the serve-job PR) replaces it with the served endpoint. Tokens are priced 0 (the
# server is paid as GPU time). `backend: local` (the default) is the HF Jobs shape; the run needs
# the task image. `swarmlab validate` and `swarmlab prompts` work offline.
name: programbench-team-piv
seeds: [0]
budget: {hard_usd: 5.0}
providers:
served: {type: openai_compat, params: {name: served, base_url_env: SWARMLAB_ENDPOINT_URL, history_reasoning: true,
api_key_env: HF_TOKEN, timeout_s: 900, pricing: {flint: [0, 0, 0]}, self_hosted: true}}
options: {commit: async, rounds_per_agent: 120, max_calls_per_turn: 40}
roles:
lead: {tools: [bash, post, read_board, wait],
channels_write: [worker-1, worker-2, worker-3, worker-4]}
worker: {tools: [bash, post, read_board, wait],
channels_read: [worker-1, worker-2, worker-3, worker-4]}
arms:
team-piv:
world: {type: programbench, params: {task: abishekvashok__cmatrix.5c082c6}}
medium:
delivery: pull
topology:
type: rooms
params:
rooms:
lead: [a000, a001, a002, a003, a004]
worker-1: [a000, a001]
worker-2: [a000, a002]
worker-3: [a000, a003]
worker-4: [a000, a004]
participants:
- {type: mini_swe_agent, count: 1, role: lead,
params: {model: "served:flint", templates: "file:/data/workspaces/swarmlab/experiments/flint-programbench/prompts/lead-template-team.yaml", steps_per_turn: 10, step_limit: 1000,
max_tokens: 32768, wait_tool: true, guidance: "file:/data/workspaces/swarmlab/experiments/flint-programbench/prompts/piv-lead.txt"}}
- {type: mini_swe_agent, count: 4, role: worker,
params: {model: "served:flint", templates: worker, steps_per_turn: 10, step_limit: 300,
max_tokens: 32768, idle_until_assigned: true, guidance: "file:/data/workspaces/swarmlab/experiments/flint-programbench/prompts/piv-worker.txt"}}

Xet Storage Details

Size:
3.74 kB
·
Xet hash:
6e9b191cf5c8476d5b4a130553a672cab439e27a09b1e58d58fc1ad5359f280d

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.