Skip to content

v2.10.4

Latest

Choose a tag to compare

@github-actions github-actions released this 20 Nov 07:04
50e66d2

Added

  • RL training command group: New bl finetune rl group for end-to-end reinforcement learning on Bailian.
    • bl finetune rl authorize checks the admission whitelist, service-linked roles and service activation. Roles are never created automatically; a missing role stops with bilingual manual-authorization guidance. --authorize enables the RL training mode (high-risk, requires --yes).
    • bl finetune rl run submits jobs from flags or a YAML spec and packages Rollout/Reward function code. Submission executes inside a managed, OS-sandboxed Python runtime (macOS/Linux only). Creating a job consumes billed MTU capacity and is gated by dry-run → explicit confirmation → --yes; --watch polls to a terminal state and a timeout does not cancel the cloud job.
    • bl finetune rl get / list / watch / logs / cancel / delete manage jobs through the shared fine-tunes backend; get returns the full OpenAPI response including datasets and MTU spec/capacity.
  • OPD distillation command group: New bl finetune opd group for pure-model online distillation (pg_opd) from a platform-registered teacher model.
    • bl finetune opd authorize / run / params / get / watch / logs / cancel / delete; run previews the job spec with --dry-run before the billed submit, and params queries supported hyper-parameters for a student base model.
  • Managed runtime commands: New bl finetune runtime prepare / doctor / clean for the RL/OPD sandbox runtime.
    • prepare downloads a pinned CPython + dashscope runtime from reviewed multi-source manifests with SHA256 verification and mirror fallback, then caches it for reuse; BAILIAN_RL_PYTHON_MIRROR / BAILIAN_RL_PIP_INDEX_URL cover restricted regions and BAILIAN_RL_RUNTIME reuses a local development runtime.
    • doctor verifies sandbox isolation (credential read denial, home redirection, SDK import, proxy egress control) and probes host outbound network reachability as an advisory check, so agent sandboxes that block network are reported with guidance instead of failing mid-submission.
    • clean reclaims orphaned task directories and superseded runtime versions with a --dry-run preview; the active baseline runtime is never touched.
  • Sandbox network-denial hints: Connect-level network denials (EPERM / EACCES / ENETUNREACH / EHOSTUNREACH) now produce bilingual hints explaining that the execution environment (for example an AI agent sandbox) may be blocking outbound network, with guidance to enable network access, run in a regular terminal, or set HTTPS_PROXY. Runtime download failures also carry the underlying errno.

Changed

  • Proxy environment variables: Lowercase http_proxy / https_proxy / no_proxy are now honored as a fallback when the uppercase variants are unset; uppercase still takes precedence.