Xnizzorg/pondllm-qwen3.5-0.8b-base-communication-sft

PondLLM Qwen3.5 0.8B communication adapter

This is a local experimental LoRA adapter for Qwen/Qwen3.5-0.8B-Base. It produces bounded JSON actions for the PondLLM simulator. It is not a general chat model.

The V3 curriculum retains the balanced action examples and adds paired communication cases:

• —signal a locally visible food coordinate when a nearby organism cannot see it;
• —suppress a redundant signal when the recipient already sees the food;
• —move toward food described in received memory;
• —preserve the control behavior when that memory is absent.

Training

• —Records: 20,006 (19,806 train, 200 evaluation)
• —Method: 4-bit QLoRA SFT
• —LoRA: rank 8, alpha 16, dropout 0.05
• —Maximum length: 1,024
• —Effective batch size: 32
• —Learning rate: 2e-4
• —Epochs: 2
• —Seed: 7
• —Runtime on RTX 5070 Ti: 7,383.8 seconds
• —Aggregate train loss: 0.05906

Exact parameters and framework versions are in run_manifest.json.

Evaluation

On 600 disjoint deterministic communication cases:

• —syntax valid: 100.00%
• —legal actions: 100.00%
• —useful-signal rate: 96.67%
• —redundant-signal rate: 0.00%
• —payload-coordinate accuracy: 96.67%
• —recipient-informed exact accuracy: 99.33%
• —recipient-control exact accuracy: 100.00%

On the older 600-case fixed-stratified action set, syntax was 100.00%, legality 99.83%, macro action accuracy 66.33%, and exact accuracy 51.83%. The four common deterministic actions were preserved or improved relative to the prior 0.8B adapter, while sharing weakened.

See the night-four communication report for comparisons, limitations, hashes, and live-world results.

Limitations

This adapter imitates a specified communication policy. It does not demonstrate emergent language, subjective experience, or improved evolutionary fitness. Its JSON schema is specific to PondLLM. The base model is required to use the adapter.