Xnizzorg/pondllm-qwen3.5-0.8b-base-v4-simulator-native

PondLLM Qwen3.5 0.8B Base action adapter V4

This is the simulator-native V4 LoRA action-policy adapter for PondLLM. It maps an organism's local observation and lifetime memory to one strict JSON action. The frozen shared base model is Qwen/Qwen3.5-0.8B-Base at revision dc7cdfe2ee4154fa7e30f5b51ca41bfa40174e68.

V4 was trained from real deterministic simulator observations and transitions. Its curriculum targets coordinate signalling, information-following, currently visible food overrides, safe resource sharing, and low-energy food priority without role-shaped IDs or lineage cues.

Evaluation

Greedy held-out evaluation produced:

• —100% strict JSON and 100% observation-legal actions;
• —98% useful signalling, 1% redundant signalling, and 98% correct signal payloads;
• —98% informed-recipient accuracy and 99.5% visible-food override accuracy;
• —100% safe-share accuracy, 0.4% unsafe-reserve sharing, and no unnecessary sharing;
• —99.7% exact accuracy on the 1,000-example survival counterfactual suite.

In 16 matched simulator layouts, the recipient foraged in 93.75% of normal-channel worlds, 0% with delivery blocked, and 6.25% with corrupted coordinates. The normal-minus-blocked causal effect was +93.75 percentage points.

The complete methodology, gates, baselines, hashes, limitations, and commands are in the V4 results report.

Use

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_id = "Qwen/Qwen3.5-0.8B-Base"
adapter_id = "Xnizzorg/pondllm-qwen3.5-0.8b-base-v4-simulator-native"

tokenizer = AutoTokenizer.from_pretrained(base_id)
base = AutoModelForCausalLM.from_pretrained(base_id, device_map="auto")
model = PeftModel.from_pretrained(base, adapter_id)

Use PondLLM's prompt builder and strict action parser rather than treating this as a general chat adapter. Generation evaluation used greedy decoding.

Training

The adapter was trained fresh for two epochs on 27,014 records with seed 7, 4-bit NF4 base-model loading, BF16 compute, completion-only loss, LoRA rank 8 and alpha 16 over all linear layers, learning rate 2e-4, micro-batch size 16, and gradient accumulation 2. Training took 10,399 seconds on an NVIDIA GeForce RTX 5070 Ti.

Dataset SHA-256: 9BDE22C42DC62B66B26449857C8005E41CE5ABEBF9BEA6C16ED72467CBAEF7EE

Adapter SHA-256: E0B76547781DB57F94D19175A9791430362BC9B5355EA7B221F67B5C4477E82B

Limitations

This is a narrow research action policy, not a general-purpose language model. Controlled two-agent communication succeeds, but an ordinary four-seed probe found no spontaneous signals in 13 richer multi-agent opportunities. Two low-energy offspring also pursued food they could not reach before exhaustion. These are explicit V4.1 targets. Results cover short fixed-seed simulations and do not establish population fitness, open-ended evolution, or emergent language.

This is supervised adapter training only. Population-scale reinforcement learning remains intentionally deferred.