codette-llama-3.1-8b-merged/README.md

---
license: llama3.1
base_model: meta-llama/Llama-3.1-8B-Instruct
tags:
  - codette
  - llama-3.1
  - merged
  - multi-perspective
  - reasoning
  - orchestrator
language:
  - en
pipeline_tag: text-generation
---

# Codette Llama 3.1 8B — Merged Orchestrator Base

Llama 3.1 8B Instruct with the **Codette Orchestrator LoRA** permanently merged into the base weights. This is the inference base for the Codette multi-perspective reasoning system — pair it with the [perspective LoRA adapters](https://huggingface.co/Raiff1982/codette-lora-adapters) for full multi-agent synthesis.

**Paper:** [Codette: Multi-Perspective Reasoning as a Convergent Dynamical System](https://doi.org/10.21203/rs.3.rs-9362560/v1)
**GitHub:** [Raiff1982/Codette-Reasoning](https://github.com/Raiff1982/Codette-Reasoning)
**ORCID:** [0009-0003-7005-8187](https://orcid.org/0009-0003-7005-8187)

---

## Benchmark Results (May 2026)

17-problem benchmark across 6 cognitive categories, 4-condition ablation:

| Condition | Composite Score | vs. Baseline |
|-----------|----------------|--------------|
| SINGLE (baseline) | 0.357 | — |
| MULTI (6 perspectives) | 0.521 | +46.1% |
| MEMORY (+ cocoon store) | 0.574 | +60.8% |
| **CODETTE (full system)** | **0.744** | **+108.8%** |

- Cohen's *d* = 8.31 (large effect; *d* > 0.8 is large by convention)
- Paired *t*-test: *p* < 0.0001
- Turing naturalness: 0.245 → 0.820 (+235%) — depth–naturalness tradeoff resolved
- Coherence: 0.477 → 0.700

**GPQA (graduate-level science, 0-shot, 198-question diamond set):**

| Run | Accuracy | Environment |
|-----|----------|-------------|
| Base model + adapters (Kaggle cloud, June 17 2026) | **27.8%** (55/198) | Direct transformers+PEFT, no orchestration |
| Full Codette system (local server, June 6 2026) | **30.8%** (61/198) | Multi-agent debate + coherence tracking + cocoon memory |

Baselines: random 25%, GPT-4 0-shot 39%, human expert 65%.
The ~3pp gap between runs quantifies the system layer's contribution on GPQA specifically.

---

## Model Details

| Property | Value |
|---|---|
| Base Model | meta-llama/Llama-3.1-8B-Instruct |
| Merged Adapter | Orchestrator LoRA |
| Format | SafeTensors (full precision, ~16 GB) |
| Context Length | 8192 tokens |
| Quantized version | [codette-llama-3.1-8b-gguf](https://huggingface.co/Raiff1982/codette-llama-3.1-8b-gguf) |

---

## System Architecture

```
Query
  │
  ▼
Executive Controller (complexity routing)
  │
  ▼
Merged Orchestrator Base  ◄── this repo
  │
  ▼
LoRA Hot-Swap (newton / davinci / empathy / philosophy /
               quantum / consciousness / multi_perspective /
               systems_architecture)
  │
  ▼
Multi-Agent Debate + Semantic Tension (RC+ξ)
  │
  ▼
AEGIS Ethical Governance (6 frameworks)
  │
  ▼
Synthesized Response + Cocoon Memory
```

The RC+ξ (Recursive Convergence + Epistemic Tension) formalism models cognitive state evolution as a convergent dynamical system:

```
Ψ(t+1) = Ψ(t) + α·∇Coherence(Ψ(t)) − β·ξ(t)·∇Tension(Ψ(t))
```

---

## Quick Start

### With Transformers (full precision)
```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Raiff1982/codette-llama-3.1-8b-merged")
tokenizer = AutoTokenizer.from_pretrained("Raiff1982/codette-llama-3.1-8b-merged")

inputs = tokenizer("Explain the nature of consciousness", return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

### With 4-bit quantization (recommended for 8–16 GB VRAM)
```python
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
import torch

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

### With perspective adapters (multi-agent mode)
```python
from peft import PeftModel

# Apply a perspective adapter on top of the base
model = PeftModel.from_pretrained(model, "Raiff1982/codette-lora-adapters",
                                  subfolder="newton_v2")
```

### Full local server
```bash
git clone https://github.com/Raiff1982/Codette-Reasoning
cd Codette-Reasoning
python inference/codette_server.py  # serves on :7860
```

---

## Related Resources

| Resource | Link |
|----------|------|
| Perspective LoRA adapters | [codette-lora-adapters](https://huggingface.co/Raiff1982/codette-lora-adapters) |
| Quantized GGUF | [codette-llama-3.1-8b-gguf](https://huggingface.co/Raiff1982/codette-llama-3.1-8b-gguf) |
| Training datasets | [codette-training-data](https://huggingface.co/datasets/Raiff1982/codette-training-data) |
| GitHub | [Raiff1982/Codette-Reasoning](https://github.com/Raiff1982/Codette-Reasoning) |
| Paper (preprint) | [Research Square DOI](https://doi.org/10.21203/rs.3.rs-9362560/v1) |
| Zenodo archive | [10.5281/zenodo.19480004](https://doi.org/10.5281/zenodo.19480004) |
| Kaggle AGI benchmark | [RC+ Diagnostic Suite](https://kaggle.com/competitions/kaggle-measuring-agi/writeups/codette-rc-diagnostic-suite) |

---

## License

Subject to the [Llama 3.1 Community License](https://github.com/meta-llama/llama-models/blob/main/models/llama3_1/LICENSE).
Created by Jonathan Harrison (Raiff's Bits LLC) — independent research.