Commits

Commits on Feb 18, 2026

fix(ml_log): use resume="allow" instead of resume="must" for wandb

neevparikh
and
claude
committed
c2eb949
feat(ml_log): resume wandb run on checkpoint resume

neevparikh
and
claude
committed
b0f0618

Commits on Feb 11, 2026

feat: update to viewer and other things
neevparikh
committed
105293e
feat: update to viewer and other things
neevparikh
committed
4eeeefb
fix(rollout_viewer): surface tool calls and tool result metadata

neevparikh
and
claude
committed
222e661

Commits on Feb 10, 2026

fix(gpt_oss): merge system messages per Harmony spec

neevparikh
and
claude
committed
c78007f

Commits on Feb 9, 2026

fix(rl): catch worker exceptions to prevent training deadlocks

neevparikh
and
claude
committed
dddb93f

Commits on Feb 7, 2026

fix(gpt_oss): map _gptoss_internal_system role to system in to_openai_message

neevparikh
and
claude
committed
a160767

Commits on Feb 6, 2026

fix: inline constant-reward filtering to keep builders/groups aligned

neevparikh
and
claude
committed
aeefebd
fix: widen rollout error catch to APIError (includes connection/timeout)

neevparikh
and
claude
committed
88d6a40
fix: filter failed rollout groups in sync training path

neevparikh
and
claude
committed
bcf2f4e
fix: filter failed rollout groups in sync training path

neevparikh
and
claude
committed
7f03d08
fix: catch APIStatusError in RL rollout worker to prevent training hang

neevparikh
and
claude
committed
0aa8082
fix: improve logging and handling for constant-reward trajectory groups

neevparikh
and
claude
committed
8694dca
chore: add py.typed marker for PEP 561 compliance

neevparikh
and
claude
committed
dcaffc8
feat: display stop_reason in rollout viewer TUI

neevparikh
and
claude
committed
c0d77fd
fix: enable scrolling on rollout viewer sidebar

neevparikh
and
claude
committed
8a9260a
feat: add scrolling to tui
neevparikh
committed
573132f
feat: Expose sampling metadata (stop_reason) through RL pipeline (#1)

neevparikh
and
claude
committed
baa8c77
[Feature] Add library for training tool-use agents (#311)

tyler-griggs
and
claude
authored
d643514

Commits on Feb 5, 2026

Remove ToolCallPart/UnparsedToolCallPart from ContentPart (#353)

joschu
and
claude
authored
b199f2e
Kimi K2.5 (Text Only) (#352)
erikwijmans
authored
f465034
Support structured content in ConversationFormatter (#343)

joschu
and
claude
authored
13b38d8
Fix XSS vulnerability in logtree and add HTML explanation (#337)

joschu
and
claude
authored
ad69942
Add temperature parameter to TinkerMessageCompleter (#336)

joschu
and
claude
authored
435efb9
Strip EOS token in RoleColonRenderer.parse_response (#339)

joschu
and
claude
authored
259500c
[tinker-cookbook] add a way to register custom renderers & tokenizers (#349)
kennyyu
authored
6888a4e

Commits on Feb 3, 2026

Add `build_supervised_examples` and `LAST_ASSISTANT_TURN`. Fix Kimi K2 thinking preservation in renderer (#341)
TonyLianLong
authored
1cfd252

Commits on Feb 2, 2026

typos + wording standardization (#318)
josancamon19
authored
44509bd

Commits on Jan 31, 2026

Add streaming parsing for the Kimi K2 renderer (#319)

nealwu
and
claude
authored
dee2fe4

Commits on Jan 30, 2026

Reuse KL reference client instead of recreating per minibatch (#332)
jeffrey-tml
authored
abea3ee

Commits on Jan 29, 2026

Changelog update 2026-01-29 (#329)

danobi
and
claude
authored
49aeed4

Commits on Jan 28, 2026

Add max_step parameter to on-policy distillation (#328)
ZitongYang
authored
c94af70

Commits on Jan 27, 2026

Add configurable KL penalty reference model (#326)
jeffrey-tml
authored
2575914
Set 7 day TTL by default on checkpoints (#324)
danobi
authored
dccee28