darioooooo0o/tiny-agent-qwen3.8-trajectories

Part of a fun research project: training a small model from scratch on one home GPU as fast as possible, for about €4–5 of electricity. These trajectories were generated in ~10 hours on that same GPU. They cover narrow synthetic tasks only. By @imdariotoo

tiny-agent Qwen3.8-27B trajectories

11,916 agent episodes by Qwen3.8-27B (GPTQ-int4, vLLM, thinking off), each solving a task in a synthetic sandbox project with five tools: bash, read, edit, write and submit. 9,310 are usable (correct, grounded, well-formed, no repeated calls, fit in 4k tokens). They were used to distill tiny-agent-112m.

Tasks

Each episode gets a freshly generated fake project: configs in yaml/json/toml/env, Python modules with tests, docs, logs, CSVs, and a long job list. All names and values are invented, so the answer exists only in the files. There are 12 task kinds:

• Lookups: config value, code constant, CSV row, doc fact, job in a long file, where a function is defined.
• Computing and counting: run a function, count services per region, sum a CSV column.
• Changes: edit a config, fix a failing test.
• Refusal: answer NOT_FOUND when the thing doesn't exist.

The generator and checker are code/tiny_agent/tasks.py in the model repo. Any task can be regenerated exactly from its seed and kind: make_task(random.Random(seed), kind).

Fields

Field
seed, kind	Regenerate the task (and its sandbox files) exactly
messages	System prompt, question, then alternating assistant turns (think, content, tool_calls) and tool results
correct / grounded	Checked by the task's verifier / answer appears in a tool result
usable, reject_reason	null, or one of wrong, bad (malformed tool call), repeat, long
prompt_format	json or qwen_native_xml (see below)
submitted, answer, calls, completion_tokens

How it was generated

• Thinking was switched off (enable_thinking=false). Qwen was asked to write one or two sentences of reasoning before each call, stored as the turn's think, at ~50 tokens per call.
• The first 1,357 episodes (prompt_format: json) asked Qwen for JSON tool calls. Qwen3.8 kept blending those with its native XML call format, so only 267 of them are usable.
• The rest (qwen_native_xml) used Qwen's own tool preamble and XML format, converted to JSON calls afterwards. That raised the usable rate to ~85%.
• The two hardest kinds (counting across configs, NOT_FOUND) were sampled twice as often after the first hour.

prompt_format	usable	wrong	bad	long	repeat
json	267	749	341	0	0
qwen_native_xml	9,043	1,200	0	315	1

License

Apache-2.0, the same as the teacher model's license. The task content is synthetic, generated by our own code.