Skip to content

README.md

Latest commit

 

History

History
60 lines (42 loc) · 1.68 KB

README.md

File metadata and controls

60 lines (42 loc) · 1.68 KB
Raw

configs

Starter TOMLs for v1 eval, legacy GEPA, and hosted training examples.

Layout

Path Used by Shape
Numbered directories Guide-specific commands v1 eval or training configs used in each guide.
eval/ uv run eval @ <file> One v1 taskset per file, tuned for a model family.
rl/ Hosted training / prime-rl configs Training configs embedding v1 env definitions under [[orchestrator.train.env]].
gepa/ prime gepa run <file> Legacy v0 GEPA configs.
endpoints.toml Prime CLI helpers Endpoint aliases for commands that still read the shared registry.

v1 Eval Configs

A v1 eval config selects one taskset and, optionally, one harness:

model = "openai/gpt-5.4-nano"
num_tasks = 20
num_rollouts = 1
max_turns = 6

[sampling]
max_tokens = 1024

[taskset]
id = "wordle"
num_tasks = 100

[harness]
id = "default"

Run it with:

uv run eval @ configs/04/wordle-eval.toml

Taskset-owned fields go under [taskset]. Harness-owned fields go under [harness]. Tool and user configs nest under the taskset field that owns them, for example [taskset.tools].

Training Configs

Training configs embed the same v1 env definition:

[[orchestrator.train.env]]
name = "wiki-search"
max_turns = 8
taskset = { id = "wiki-search", max_examples = 512, tools = { shared = true } }
harness = { id = "default" }

Add another [[orchestrator.train.env]] block only when you deliberately train on multiple environments.

GEPA Configs

The GEPA CLI in this checkout still loads v0 environments. Files under configs/gepa/ are kept for that legacy workflow and should not be used as v1 taskset config examples.