Skip to content

SeisoLocalAI v0.3.0

Choose a tag to compare

View all tags
@Legendarylibr Legendarylibr released this 21 Jul 05:23
· 265 commits to main since this release
v0.3.0
9598355
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

SeisoLocalAI v0.3.0

Since v0.2.0 (2026-07-15). ~130 commits focused on RL/SFT correctness, real verifiable data paths, Forge safety, and kernel/chat quality.

Highlights

RL / GRPO / DPO correctness

  • Outcome-first rewards: refuse format ties, field/process reward hacks, zero-spread groups, and stale vLLM/SGLang weight sync (#385).
  • DeepSeek / OpenRLHF / TRL-aligned math: dual-clip, seq-mean loss, truncated/empty excluded from advantages, DPO log 2 identity and empty-completion refusal (#385).
  • Code GRPO binary/dense/auto outcomes; contains-answer hardening; held-out slime eval (#367–#370).
  • Real-data gates for slime / Distill-RL (Forge UI + product floors) (#371).

Safety & platform

  • PENDING job cancel; fail-closed remote code-exec; DNS pin without global getaddrinfo (#374, #377, #384).
  • RL-quant deploy claims require llama.cpp + external quality sidecar (#375).
  • Refuse FULL/BASE export of LoRA-only checkpoints; restore inference kernels on load failure (#379, #380).
  • Torch 2.13 (CVE-2025-3000); private GitHub security advisory policy (#355–#356).

Training & chat

  • Unified SFT tokenization / structure-preserving prep; packing + response-mask guards (#364–#366).
  • Thinking budgets / strip leaked think blocks; empty Qwen3/Ollama reply fixes (#359–#360).
  • Kernel stack: fused GEMM/LoRA routing, attention/packing defaults (#358–#359).

Cleanup

  • Remove RL-quant markdown research-report stage; unused Forge DB tables; dead helpers (#372–#373, #382).
  • July 2026 codebase review report (#376).

Upgrade notes

  • Slime HTTP rollouts require *_sync_weights: true (debug-only env escape).
  • Default auto_stop_metric is outcome_reward_mean; loss_aggregation defaults to seq_mean; clip_ratio_c defaults to 3.0.
  • Distill-RL requires rollout_temperature > 0 and grpo_group_size >= 2.
  • Product slime / Distill-RL paths expect real verifiable data (CI uses SEISO_ALLOW_TINY_RL=1).

Full changelog: v0.2.0...v0.3.0

Assets 2
Loading