HerrHruby/Qwen3.5-4B-TMax-CISPO

Qwen3.5-4B-TMax-CISPO

Qwen3.5-4B fine-tuned with CISPO (Clipped IS-weight Policy Optimization) on the TMax-15K terminal-agent RL environment, using a fully-asynchronous rollout/trainer setup (verl).

Training

• —Base model: Qwen/Qwen3.5-4B
• —Algorithm: CISPO (rollout-anchored), clip high 0.28 / low 10
• —Sampling: temperature 1.0, top_p 1.0, group size 16
• —Data: TMax-15K, text-only short/moderate complexity split (AppTainer-compatible allowlist)
• —Agent: terminal_echo_tool_agent (Terminus-2 command interface)
• —Precision: fp32 generation/LM head + fused chunked cross-entropy
• —Exported from trainer checkpoint (global_step 51).

Intended use

Research checkpoint for terminal/agentic RL. Not instruction-tuned for general chat.