Codette LoRA Adapters — 10 Perspective Lenses

10 specialized LoRA adapters for the Codette Multi-Perspective Reasoning System. Each adapter encodes a distinct cognitive stance; the system routes queries and synthesizes across them at inference time.

Paper: Codette: Multi-Perspective Reasoning as a Convergent Dynamical System Base model: Raiff1982/codette-llama-3.1-8b-merged

Behavioral Verification (June 17, 2026)

All 7 perspective adapters passed behavioral verification after v2 retrain — zero template-contamination markers detected across 14 targeted probes. Adapters generate substantive, specific responses rather than training-data boilerplate.

Adapters

Adapter	Cognitive Stance	Version	GGUF
newton	Analytical physics, mathematical precision, conservation laws	v2	`newton-behavioral-lora-f16.gguf`
davinci	Creative invention, cross-domain synthesis, analogical reasoning	v2	`davinci-behavioral-lora-f16.gguf`
empathy	Emotional intelligence, active listening, anti-flattery	v2	`empathy-behavioral-lora-f16.gguf`
philosophy	Epistemological analysis, Socratic method, ethical frameworks	v2	`philosophy-behavioral-lora-f16.gguf`
quantum	Probabilistic reasoning, Bayesian updating, superposition thinking	v2	`quantum-behavioral-lora-f16.gguf`
consciousness	Recursive meta-cognition, RC+ξ self-monitoring	v2	`consciousness-behavioral-lora-f16.gguf`
multi_perspective	Cross-lens synthesis, integrative reasoning	v2	`multi_perspective-behavioral-lora-f16.gguf`
systems_architecture	Modularity, scalability, engineering tradeoffs	v2	`systems_architecture-behavioral-lora-f16.gguf`
orchestrator	Query routing, debate coordination, coherence monitoring	v1	`orchestrator-behavioral-lora-f16.gguf`
constraint_tracker	Instruction compliance, constraint enforcement	v1	`constraint_tracker-behavioral-lora-f16.gguf`

v2 Adapter Training (June 15, 2026)

Hand-authored datasets replacing the original template-generated training data
Jargon-free system prompts (removed RC+ξ metric references that encouraged filler)
6 epochs, lr 1e-4, rank 16, A10G GPU
Science claims web-verified before inclusion in training data
Available as PEFT safetensors under {name}_v2/ subfolders

Usage

Hot-swap with llama-cpp-python (local inference)

from llama_cpp import Llama

llm = Llama(model_path="Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
            n_ctx=8192, n_gpu_layers=35)
llm.load_lora("newton-behavioral-lora-f16.gguf")

response = llm.create_chat_completion(
    messages=[{"role": "user", "content": "Derive the escape velocity formula."}],
    max_tokens=512,
)

With PEFT + transformers (fine-tuning / Kaggle)

from transformers import AutoModelForCausalLM, BitsAndBytesConfig
from peft import PeftModel
import torch

bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4",
                         bnb_4bit_compute_dtype=torch.bfloat16)
base = AutoModelForCausalLM.from_pretrained(
    "Raiff1982/codette-llama-3.1-8b-merged",
    quantization_config=bnb, device_map="auto"
)
model = PeftModel.from_pretrained(base, "Raiff1982/codette-lora-adapters",
                                  subfolder="newton_v2")

Full Codette server (all adapters, live routing)

git clone https://github.com/Raiff1982/Codette-Reasoning
cd Codette-Reasoning
python inference/codette_server.py  # :7860

File Structure

codette-lora-adapters/
  ├── behavioral-gguf/              # v2 GGUF adapters (27 MB each)
  │   ├── newton-behavioral-lora-f16.gguf
  │   ├── davinci-behavioral-lora-f16.gguf
  │   ├── empathy-behavioral-lora-f16.gguf
  │   └── ...
  ├── newton_v2/                    # v2 PEFT safetensors
  │   ├── adapter_config.json
  │   └── adapter_model.safetensors
  ├── davinci_v2/
  ├── empathy_v2/
  ├── ...
  ├── newton/                       # v1 PEFT safetensors (legacy)
  └── ...

Related Resources

Resource	Link
Merged base model	codette-llama-3.1-8b-merged
Quantized GGUF base	codette-llama-3.1-8b-gguf
Training datasets	codette-training-data
GitHub	Raiff1982/Codette-Reasoning
Paper	Research Square DOI