Views
No views yet
environment/agent conversation roles,
Qwen-XML tool calls after </think>. Trained on the v2 no-run mix (140,730 rows across
agentic loops, plan-act handoff, RDDL planning, chat, instruct, creative writing;
sources incl. TextArena, Factorio, Diplomacy, Jericho, ScienceWorld, xlam, gsm8k-grounded).<run> blocks
were excluded (uses_run=False filter) to isolate pipeline validation from the REPL
architecture change. The REPL-enabled sibling trains on the full mix.chat_template.jinja) — it accepts roles
system/environment/agent/user and will raise on anything else.<|im_start|>agent\n. Stop on <|im_end|> (id 248046).<tool_call><function=NAME><parameter=KEY>...</parameter></function></tool_call>