artivus-ai/vulcan-agent-0.8b-convergence-sft-v2

Vulcan-Agent 0.8B — Convergence SFT v2 (no-run shakeout)

First checkpoint lineage in the convergence turn format: structured thinking block (REFLECTION/STATE/GOAL/PLAN/PREDICT/DECISION), 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).

This is the EfficientZero-shakeout variant: rows containing executable <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.

Usage notes

• Chat template ships in the repo (chat_template.jinja) — it accepts roles system/environment/agent/user and will raise on anything else.
• Generation prompt: <|im_start|>agent\n. Stop on <|im_end|> (id 248046).
• Tool calls: <tool_call><function=NAME><parameter=KEY>...</parameter></function></tool_call>
• Part of the Vulcan cognitive-core program (Artivus).