# Copy this file to config.toml. Generated data and runs are gitignored.

[dataset]
# Downloaded from https://www.kaggle.com/datasets/dheerajmpai/projecteuler
archive = "~/Downloads/euler-questions.zip"
prepared_dir = "data"

[model]
pi_command = "pi"
# Examples:
# provider = "tinker"
# model = "thinkingmachines/Inkling-Small:peft:262144"
# provider = "zai"
# model = "glm-5.3"
provider = "tinker"
model = "thinkingmachines/Inkling-Small:peft:262144"
thinking = "high"
# Set this to a checkpoint alias/hash when a proxy model ID stays constant.
revision = ""

[evaluation]
languages = ["python", "javascript", "julia"]
strategies = ["derive", "bruteforce_then_optimize", "tests_first", "recurrence", "readable"]
# Use 2 for a cheap smoke test. A serious gate should use at least 8-16.
fanout_per_language = 2
max_concurrency = 2
timeout_seconds = 600

[sandbox]
# Pi's orchestration process remains on the host. Its read/write/edit/bash effects
# are workspace-confined and execute through this networkless Docker container.
enabled = true
image = "volcengine/sandbox-fusion@sha256:dd7ff53d16132a8acad6d5da7f15154bb4a331381567a4cb21b3e97ce581f5f9"
platform = "linux/amd64"
memory = "768m"
cpus = "1.0"
pids_limit = 64

[curriculum]
start_problem = 1
state_file = "data/curriculum-state.json"
# For fanout=2 this keeps the example runnable. Raise to 6 with fanout=8.
required_correct_per_language = 1
required_consecutive_rounds = 2
regression_window = 3
minimum_regression_rate = 0.8
require_successful_execution = true
# functional_v2 requires a successful write/edit, zero-exit execution of that same
# relative file, exact output, a semantically correct response, no escape attempt,
# clean completion, and Docker isolation.
qualification_mode = "functional_v2"