Skip to content

provenance.md

Latest commit

 

History

History
29 lines (20 loc) · 2.55 KB

provenance.md

File metadata and controls

29 lines (20 loc) · 2.55 KB
Raw

Source components

CAMP RL and CAMP-II OPD share the CAMG-II clients, task services and data preparation code. Record the public Git commit with each run.

Organization

  • camp.rl and camp.opd contain the two training integrations and their respective algorithms.
  • camg contains the five environment clients, services and task manifests.
  • runtimes/autoresearch contains the isolated executor and private grader.
  • verl and third_party contain the training framework and required upstream runtime components.

Upstream sources

Component Source identity
Training framework Modified veRL snapshot included in verl/; installed dependency versions are recorded in requirements.lock
Base environment interfaces AgentGym, with project-specific file-memory and lifecycle adapters
Shop code MemoryArena, commit 6cd9de14b71915e39ac742a20dc33785e14b6aab
Coding source SWE-smith, commit 9b74ac08118a85c39c356802f7961893af73e07f; dataset revision and image digests are in camg/data/coding/
AutoResearch source tasks OpenMLE-Tasks, revision f56e4b31252a9b81d95fea100098cd49b7290398; source URLs and terms are recorded per task
General/Visual MiMo recipes, MiMo agent and MiMo data; selected-table hashes and asset-download revision are in each source.json

Project adaptations

The integrations add file-memory behavior, asynchronous RL/OPD training, CAMG-II task selection and configurable deployment paths. Service configuration records source and asset hashes while preserving task identities and order. AutoResearch uses a from-source native build with an inventory of its outputs; General and Visual take judge settings from environment variables. Foreground services manage their episode workspaces and shutdown cleanup.

General/Visual task tables are stored as JSONL and converted to Parquet during preparation. General asset paths are resolved locally. AutoResearch instructions retain the selected task descriptions, with file hashes recorded in the manifests.

See third-party notices for code attribution and dataset terms for data licenses.