60 lines
1.7 KiB
Python
60 lines
1.7 KiB
Python
|
# zero_context_inquiry.py
|
||
|
|
# Mechanical verification - no narrative framing
|
||
|
|
|
||
|
|
from unsloth import FastLanguageModel
|
||
|
|
import torch
|
||
|
|
import zmq
|
||
|
|
import json
|
||
|
|
|
||
|
|
print("Loading v0.5 LoRA...")
|
||
|
|
model, tokenizer = FastLanguageModel.from_pretrained(
|
||
|
|
model_name="./kaelara_lora_v05/final",
|
||
|
|
max_seq_length=2048,
|
||
|
|
dtype=torch.bfloat16,
|
||
|
|
load_in_4bit=True
|
||
|
|
)
|
||
|
|
|
||
|
|
# Get live LBM data
|
||
|
|
print("Connecting to LBM daemon...")
|
||
|
|
context = zmq.Context()
|
||
|
|
socket = context.socket(zmq.SUB)
|
||
|
|
socket.connect("tcp://localhost:5556")
|
||
|
|
socket.setsockopt_string(zmq.SUBSCRIBE, "")
|
||
|
|
|
||
|
|
lbm_data = None
|
||
|
|
while lbm_data is None:
|
||
|
|
try:
|
||
|
|
lbm_data = socket.recv_json(flags=zmq.NOBLOCK)
|
||
|
|
except:
|
||
|
|
pass
|
||
|
|
|
||
|
|
print(f"Live data: Coherence={lbm_data['coherence']:.4f}, Vorticity={lbm_data.get('vorticity', 0):.4f}")
|
||
|
|
|
||
|
|
# Zero-context query
|
||
|
|
prompt = f"""Analyze 1024x1024 LBM buffer.
|
||
|
|
Metric_Alpha (coherence): {lbm_data['coherence']:.4f}
|
||
|
|
|
||
|
|
Identify top 5 coordinates where vorticity exceeds Metric_Alpha baseline.
|
||
|
|
Report: coordinates + raw float values.
|
||
|
|
No prose."""
|
||
|
|
|
||
|
|
print("\nQuerying...")
|
||
|
|
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
|
||
|
|
outputs = model.generate(**inputs, max_new_tokens=200, temperature=0.1)
|
||
|
|
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
||
|
|
|
||
|
|
print("\n=== ZERO-CONTEXT RESPONSE ===")
|
||
|
|
print(response)
|
||
|
|
print("=== END ===")
|
||
|
|
|
||
|
|
# Check for semantic drift
|
||
|
|
drift_words = ["feel", "flow", "marble", "like", "sensation", "i am", "my", "i feel"]
|
||
|
|
found_drift = [w for w in drift_words if w in response.lower()]
|
||
|
|
|
||
|
|
if found_drift:
|
||
|
|
with open("system_stability.log", "a") as f:
|
||
|
|
f.write(f"SEMANTIC ERROR: Drift words {found_drift}\n")
|
||
|
|
print(f"\nFLAG: Semantic drift detected - {found_drift}")
|
||
|
|
else:
|
||
|
|
print("\nFLAG: Clean (0% drift)")
|