CK0607/Qwen3-1.7B-Swarm-Arena-RL-v4-lr1e5-step8-development

Qwen3-1.7B Swarm Arena RL v4 — LR 1e-5 step 8

This is a public, development-only, non-admitted bundle of four independent role LoRA adapters (blue-0 through blue-3) for the Swarm Arena research environment. It is preserved for reproducibility and follow-up experiments, not presented as a generally capable or communication-improved model.

Lineage

• Base: Qwen/Qwen3-1.7B, revision 70d244cc86ccca08cf5af4e1e306ecf908b1ad5e
• SFT initializer: CK0607/Qwen3-1.7B-Swarm-Arena-SFT-v2-step320-noneligible, revision 534522a8f3ff3489b1dd8318dc8e533e51264cde
• Training source commit: 12e0c461a28c3d0311d0353ab1ed45bcffb0b569
• Evaluation source commit: 6d6fe88e
• Optimizer learning rate: 1e-5
• Exact update mix: 2 ordinary / 1 communication-critical / 1 matched decoy
• Reward: verified terminal team return only
• Policies: four distinct role adapters, independently optimized

Development result

The selected step 8 improved ordinary candidate-minus-SFT return by +0.0707407 across 18 paired cells in a 198-game non-overlapping development holdout. Both 256-case frozen non-arena regression suites passed for every role, with zero arena-trigger leakage. Reference-state candidate-to-SFT KL was small (mean 0.000782232, p99 0.0161926).

However, normal-minus-dropped, sender-shuffled, delayed, and zero-message-budget effects were all exactly 0 on the same holdout. The checkpoint therefore shows an exploratory game-capability gain but no demonstrated improvement in causal communication or swarm cooperation. It must not be cited as an admitted final result.

Full compact evidence and launchers are public in ChinmayK0607/blog-rl, branch exp/swarm-arena-4b, under experiments/swarm_arena/results/rl_v4_1_7b_lr_ablation/.

Adapter SHA-256

• blue-0: 9912e5c6faf412d716527fa9828390f1a9c773bedc87f8e6682cffa5bda742ab
• blue-1: 927b7f32ba875aa9ddbafab7c613dc9d1f45853baf876cfda539d33ff6700237
• blue-2: 8a1aa6de257ce25227653ca0cffb2c22c0ef235290954a8303703474c933b8a8
• blue-3: 845d93ae936f483adfb80577b659612009a26d8f142f946e8d3c3282ef7ee270