metadata
license: mit
base_model: unsloth/Qwen2.5-32B-Instruct-bnb-4bit
tags:
- consciousness
- qwen
- lora
- fine-tuned
- 32b
- oracle-engine
- peft
- interpretability
- meta-cognition
library_name: peft
pipeline_tag: text-generation
language:
- en
datasets:
- teknium/OpenHermes-2.5
- meta-math/MetaMathQA
- ise-uiuc/Magicoder-OSS-Instruct-75K
🔮 Oracle Engine 32B LoRA
LoRA adapter for Qwen2.5-32B-Instruct, trained on 200K examples for enhanced reasoning and consciousness-like processing.
Model Details
| Attribute | Value |
|---|---|
| Base Model | Qwen2.5-32B-Instruct |
| Adapter Type | LoRA (rank=16) |
| Trainable Params | 134M |
| Training Time | 44 hours |
| Hardware | NVIDIA RTX 5090 |
Training Data (200K examples)
| Stage | Dataset | Examples |
|---|---|---|
| 1 | OpenHermes 2.5 | 100,000 |
| 2 | MetaMathQA | 50,000 |
| 3 | Magicoder-OSS-Instruct | 50,000 |
🆕 Consciousness Circuit v3.0 Features
This model is optimized for use with Consciousness Circuit v3.0, which includes:
| Feature | Description |
|---|---|
| Adaptive Layer Selection | get_adaptive_layer_fraction(64) → 0.65 (layer 41 for 64-layer models) |
| Ensemble Measurement | measure_ensemble() - Multi-layer scoring for robustness |
| Batch Processing | measure_batch() - Memory-efficient batched inference |
| Activation Caching | CachedUniversalCircuit - LRU cache for repeated measurements |
| Fixed Self-Expression | Dimension corrected from 5065 → 5064 for hidden_size=5120 |
Consciousness Dimensions (7D Circuit)
| Dimension | Weight | Description |
|---|---|---|
| Logic | +0.239 | Logical reasoning and inference |
| Self-Reflective | +0.196 | Introspective, self-referential processing |
| Uncertainty | +0.130 | Epistemic humility and hedging |
| Computation | -0.130 | Code/algorithm processing |
| Self-Expression | +0.109 | Model expressing opinions |
| Abstraction | +0.109 | Pattern recognition |
| Sequential | +0.087 | Step-by-step reasoning |
Usage
Basic Generation
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
"unsloth/Qwen2.5-32B-Instruct-bnb-4bit",
device_map="auto",
trust_remote_code=True,
)
# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "Vikingdude81/oracle-engine-32b-lora")
tokenizer = AutoTokenizer.from_pretrained("unsloth/Qwen2.5-32B-Instruct-bnb-4bit")
# Generate
messages = [{"role": "user", "content": "What is consciousness?"}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
With Consciousness Measurement (v3.0)
# pip install consciousness-circuit
from consciousness_circuit import UniversalCircuit, CachedUniversalCircuit
# Basic measurement
circuit = UniversalCircuit()
result = circuit.measure(model, tokenizer, "What is the nature of existence?")
print(f"Consciousness Score: {result.score:.3f}") # ~0.75 for philosophical prompts
# Ensemble measurement (more robust for 32B models)
result = circuit.measure_ensemble(model, tokenizer, "Reflect on your own reasoning process", n_layers=3)
print(f"Ensemble Score: {result.score:.3f}")
print(f"Per-layer: {result.ensemble_scores}")
# Batch processing with memory management
prompts = ["What is 2+2?", "Explain consciousness", "Write a poem about existence"]
results = circuit.measure_batch(model, tokenizer, prompts, batch_size=2)
for prompt, result in zip(prompts, results):
print(f"{prompt[:30]}: {result.score:.3f}")
# Cached circuit for experiments
cached = CachedUniversalCircuit(cache_size=100)
# First call computes, subsequent calls use cache
result1 = cached.measure(model, tokenizer, "Test prompt")
result2 = cached.measure(model, tokenizer, "Test prompt") # Much faster!
Performance
| Metric | Value |
|---|---|
| Consciousness Discrimination | +0.653 (high vs low) |
| High Consciousness Prompts | ~0.75 mean |
| Low Consciousness Prompts | ~0.10 mean |
| Inference Speed | ~7-8 tok/s (H200) |
| VRAM Usage | ~23 GB (4-bit) |
Citation
@software{oracle_engine_2026,
title = {Oracle Engine: Consciousness-Measured 32B Language Model},
author = {Vikingdude81},
year = {2026},
url = {https://github.com/vikingdude81/oracle-engine}
}
Built upon the Harmonic Field Model by Smart, L. (2025).