Quick Start

This is a debug-only end-to-end run of the agent-loop RL pipeline with the smallest possible setup: host deployment + 8 dummy samples + a dummy reward, on Qwen3-0.6B. It verifies that the full path works — data → agent loop → trajectory → reward → GRPO training — and nothing more. It produces no meaningful accuracy or training gains. For real training, use a real dataset, a real reward, and a larger model (seeAgent Reinforcement Learning). The runnable files live under examples/quick_start/.

Prerequisites

Uni-Agent installed (see Installation, “Single-Node Trial”)
A Qwen3-0.6B checkpoint available locally
Python ≥ 3.10

Components

Everything is already in the repo — no files to create:

File	Purpose
`uni_agent/reward/dummy.py`	Dummy reward: 1.0 if the agent called `submit`, else 0.0
`examples/quick_start/agent_config.yaml`	Host-deployment agent loop (bash + submit tools)
`examples/quick_start/generate_dummy_data.py`	Generates 8 dummy training samples
`examples/quick_start/runtime_env.yaml`	Ray runtime env (packages the repo, sets `PYTHONPATH`)
`examples/quick_start/train.sh`	Single-node GRPO launcher (`ray job submit`)

The agent config uses tool_parser: hermes so it works with Qwen3’s default JSON tool-call format (no chat-template change needed).

Generate data & train

cd ~/uni-agent

python examples/quick_start/generate_dummy_data.py --local-save-dir ~/data/swe_agent

ray start --head
export MODEL_PATH=$HOME/models/Qwen3-0.6B
export NGPUS_PER_NODE=1
bash examples/quick_start/train.sh

Verify

grep "STEP 1" ~/logs/agent/*/run.log       # agent loop ran
grep "reward_score" ~/logs/agent/*/run.log  # reward computed

Troubleshooting

Symptom	Fix
`response_mask must contain at least one valid token`	Check run.log for the real crash; increase `max_response_length`
`string indices must be integers, not 'str'`	`prompt` must be `list[dict]`, not a plain string
`No function call found in the response`	Set `tool_parser: hermes` in the agent config
`Cannot write struct with no child field`	`extra_info.tools_kwargs` must be non-empty
Tool install permission denied	Set `tool_install_dir: ~/.local/bin` in the agent config