Qwen3-4B-Swarm-Arena-SFT-v2

LoRA warm start for the Swarm Arena 4v4 partially observed graph-control environment. The model emits strict JSON broadcasts and actions for individual agents. This checkpoint was selected using validation behavior only, then evaluated once on the untouched SFT test split and frozen 60-case arena.

This is not a multi-agent-RL-trained model. It is the supervised protocol and mechanics warm start intended to make subsequent MARL experiments interpretable.

Selection and test results

• selected training step: 240;
• validation candidates: 8 (4 passed fixed gates);
• held-out test phase-balanced exactness: 0.9826;
• held-out action exactness: 0.9783;
• held-out broadcast exactness: 0.9869;
• held-out unsupported-broadcast rate: 0.0000.

Frozen arena results

• strict broadcast rate: 1.0000;
• strict generated-condition action rate: 1.0000;
• generated-condition mean oracle regret: 2.2250;
• generated-condition mean environment reward: 0.4200;
• generated-minus-dropped reward: 0.3858;
• paired coordination-improvement claim gate: false.

The claim gate requires 95% paired intervals to support all three: lower oracle regret than the untouched base model, generated messages beating dropped messages, and generated messages beating shuffled messages. A false result is reported as a negative or inconclusive experiment, not relaxed post hoc.

Reproducibility

• source commit: f1fd20fb;
• training run: https://wandb.ai/ChinmayK0604/swarm-arena-sft/runs/o0ioy6jt;
• environment: arena-core-v1;
• prompt: arena-v2-structured-priority;
• arena manifest: b53bfc523043ec71cc69f851d0819511c5a9f0b4f09520898f30954bbe874b29;
• deterministic generation with thinking disabled;
• learned artifact: rank-32 PEFT LoRA adapter over Qwen/Qwen3-4B-Instruct-2507.

The results/ directory contains the frozen selection decision and final test, arena, and paired-comparison summaries. Raw trajectories are logged separately as a W&B evaluation artifact.

Intended use

Use for research on structured communication, multi-agent credit assignment, and controlled MARL within the discrete Swarm Arena simulator. Do not interpret the abstract graph actions as real network-security capabilities.