All pull requests

fix(dataset): pair Anthropic tool_result blocks with their tool_use by id
#10293Zhuoxi2000 opened Oct 4, 20261 of 4 tasks
fix(patcher): resolve MP+DDP device_map memory through the active accelerator
#10292li-lizhe opened Oct 4, 2026Contributor1 of 4 tasks
fix(cli): handle --help/-h and missing subcommand in cli_main
#10291KaiyiQuan opened Oct 3, 2026
1
feat(model): support Jina reranker v3/v3.5
#10290haahy opened Oct 3, 20261 of 4 tasks
fix: handle cli help command
#10289Subodh-17 opened Oct 3, 2026
fix(cli): support top-level help
#10288haahy opened Oct 3, 20261 of 4 tasks
fix(template): handle assistant text before tool calls in ReAct templates and Seed-OSS
#10284Zhuoxi2000 opened Oct 1, 20261 of 4 tasks
fix: preserve Xing4.0 and MiniCPM5 tool argument types and whitespace
#10281taking-lying-flat opened Oct 1, 2026Contributor1 of 4 tasks
3
fix: respect Xing4.0 eager mode and reject conflicting patch options
#10280taking-lying-flat opened Oct 1, 2026Contributor1 of 4 tasks
fix(ui): use the last assistant turn when marking a toolbench observation
#10279RerankerGuo opened Oct 1, 2026Contributor
fix(dataset): key the map cache on the arguments of the map
#10278RerankerGuo opened Oct 1, 2026Contributor
fix(dataset): restore dataset_infos.json after loading a local dataset folder
#10277RerankerGuo opened Oct 1, 2026Contributor
fix(megatron): defer full-vocabulary GKD teacher logits to forward_step
#10276FOWEPJF255 opened Oct 1, 20261 of 2 tasks
2
test(template): edge cases for mixed rollout response (#10077)
#10269RerankerGuo opened Sep 30, 2026Contributor
fix(agent): preserve Qwen tool argument types and whitespace
#10267taking-lying-flat opened Sep 30, 2026Contributor1 of 4 tasks
fix(reward): follow the GRPO reward contract in MathORM
#10266Lesereingrape opened Sep 29, 2026Contributor
fix(utils): report port-scan exhaustion instead of returning a busy port
#10264Lesereingrape opened Sep 29, 2026Contributor1 of 4 tasks
fix(infer): inline a local image whose path is longer than 200 characters
#10262SashaMIT opened Sep 29, 2026
fix(template): preserve original tool-call arguments JSON (#10216)
#10261RerankerGuo opened Sep 29, 2026Contributor
fix(utils): restore ZeRO-3 hooks after generation errors
#10260fukalous opened Sep 29, 20261 of 4 tasks
fix(template): preserve assistant reasoning on Qwen3.5/3.6 tool-call turns (#10255)
#10259RerankerGuo opened Sep 29, 2026Contributor
perf(grpo): reuse rollout grouping for SP advantage updates
#10257taking-lying-flat opened Sep 29, 2026Contributor1 of 4 tasks
fix(grpo): normalize Liger DAPO across gradient accumulation
#10256taking-lying-flat opened Sep 29, 2026Contributor1 of 4 tasks
fix(agent_template): render tool call arguments as JSON in ReAct and GLM4 templates
#10254MohammadHijjawi97 opened Sep 28, 20261 of 4 tasks
fix(utils): treat a blank or non-numeric launch env var as unset
#10251Lesereingrape opened Sep 27, 2026Contributor1 of 4 tasks

Navigation Menu

All pull requests

fix(dataset): pair Anthropic tool_result blocks with their tool_use by id

fix(patcher): resolve MP+DDP device_map memory through the active accelerator

fix(cli): handle `--help`/`-h` and missing subcommand in cli_main

feat(model): support Jina reranker v3/v3.5

fix: handle cli help command

fix(cli): support top-level help

fix(template): handle assistant text before tool calls in ReAct templates and Seed-OSS

fix: preserve Xing4.0 and MiniCPM5 tool argument types and whitespace

fix: respect Xing4.0 eager mode and reject conflicting patch options

fix(ui): use the last assistant turn when marking a toolbench observation

fix(dataset): key the map cache on the arguments of the map

fix(dataset): restore dataset_infos.json after loading a local dataset folder

fix(megatron): defer full-vocabulary GKD teacher logits to forward_step

test(template): edge cases for mixed rollout response (#10077)

fix(agent): preserve Qwen tool argument types and whitespace

fix(reward): follow the GRPO reward contract in MathORM

fix(utils): report port-scan exhaustion instead of returning a busy port

fix(infer): inline a local image whose path is longer than 200 characters

fix(template): preserve original tool-call arguments JSON (#10216)

fix(utils): restore ZeRO-3 hooks after generation errors

fix(template): preserve assistant reasoning on Qwen3.5/3.6 tool-call turns (#10255)

perf(grpo): reuse rollout grouping for SP advantage updates

fix(grpo): normalize Liger DAPO across gradient accumulation

fix(agent_template): render tool call arguments as JSON in ReAct and GLM4 templates

fix(utils): treat a blank or non-numeric launch env var as unset

Pull requests