Arko007/zenyx-agentic-loops

Zenyx Agentic Loops

Multi-turn tool-use / agent-loop SFT data for small chat models (Zenyx-Mini, ~151M parameters). It targets exactly what small base models get wrong: using what the tool actually returned, recovering from errors, saying "nothing found" instead of inventing, asking and waiting for the user's answer, confirming before destructive actions, and polling a job until it is really done.

How the rows are made (and why they are trustworthy)

• A local 4B model (Spark-X2.6-4B, Q8_0 GGUF through llama.cpp, Apache-2.0) plays only the assistant.
• A small program (scripts/agentloop_env.py) plays the tools and the user. Every tool call is executed by a simulated world (virtual files, weather/time tables, calculator, unit conversion, orders/customers, e-mail/calendar, async jobs), so tool outputs are real, including errors, empty results and "still running".
• The user turns are scripted (a few hundred template variants with random slots). When the assistant asks a required question, the program answers it; when it asks for confirmation, the program says yes (about 70%) or no.
• Only rollouts that really succeed are accepted. A judge checks the world state and the transcript, e.g. the final answer must contain the values the tool returned and no number that appears in no observation; the e-mail must have been sent to the address the user gave; the file must be gone only if the user confirmed; the job must have been polled until done; the assistant must not have called a tool before confirmation.
• Rejected rollouts are kept under rejects/ with the reason (never deleted), so they can be used as negatives / DPO pairs.
• data/identity_nini/ is programmatic (no model): identity, size, creator, capabilities and limits of Zenyx-Mini ("Nini"), with exactly the facts we want it to state.

Format

Each line of data/<task>/*.jsonl:

field	meaning
messages	list of {role, content}, roles system (optional), user, assistant, tool. Assistant tool calls are only <tool_call>[{"name": ..., "arguments": {...}}]</tool_call>; results are tool messages with the raw result text.
tools	JSON string: the OpenAI-style tool schemas available in the row
domain	agentloop_<task> or identity_nini
provider	llama.cpp:Spark-X2.6-4B-Q8_0.gguf or program:templates
seed, task_info	reproducibility info (JSON string)

The format is the Zenyx chat format (zenyx100m/chat.py): the tool schemas go into the system turn, tool results come back as <|im_start|>tool turns; only assistant content is trained.

Task families

weather, time, calc, convert, read_note (list then read), error_recovery (wrong file name, transient 503, permanent permission error), empty_result (nothing found, no invention), ask_then_act (missing city / recipient / time; plus control rows where everything is given and the assistant must not ask), confirm_destructive (delete only after yes), async_job (start, poll, answer only when done), no_tool_needed, tool_unavailable (honest "I can't do that with these tools"), multi_call, dependent_calls (second call uses an id from the first result), write_note, and identity_nini.

Limits (read before using)

• Tools are simulated and the scenarios come from a fixed set of families: this teaches the loop behaviour, not real-world tool knowledge. English only.
• Assistant wording comes from a 4B model: it is accepted when the behaviour and facts are correct, so style can be plain or terse.
• Row counts grow while the generator runs; the current counts are in status/*.json.
• Scripts used (scripts/) are included so the whole pipeline can be reproduced; scripts/test_agentloop_env.py shows that a scripted correct agent is accepted on every task and wrong behaviours are rejected.

License: Apache-2.0 (data and scripts). Generator model: Spark-X2.6-4B (Apache-2.0).