Raiff1982/codette-llama-3.1-8b-merged

Codette Llama 3.1 8B Merged - Orchestrator Model

Full-precision Llama 3.1 8B Instruct with the Codette Orchestrator LoRA permanently merged into the base weights.

This model serves as the foundation of the Codette reasoning system. It has the orchestrator capabilities (query routing, debate coordination, coherence monitoring) baked into the weights, so no adapter loading is needed for core orchestration.

Model Details

Property	Value
Base Model	meta-llama/Llama-3.1-8B-Instruct
Merged Adapter	Orchestrator (4000 examples, 4 epochs)
Format	SafeTensors (full precision)
Size	~16 GB
Context Length	4096 tokens

Orchestrator Capabilities

The merged orchestrator adapter gives the model these built-in skills:

• Query Routing: Classifies queries as SIMPLE, MEDIUM, or COMPLEX
• Adapter Selection: Chooses optimal perspective adapters per query
• Multi-Agent Debate: Coordinates structured reasoning across perspectives
• Semantic Tension Tracking: Monitors epistemic tension (xi) between viewpoints
• Coherence Field: Detects reasoning collapse via Gamma metric
• Synthesis: Produces unified responses from multi-perspective debate
• AEGIS Governance: Applies 6-framework ethical validation

Usage

With Transformers

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))

Convert to GGUF

python convert_hf_to_gguf.py Raiff1982/codette-llama-3.1-8b-merged --outtype q4_k_m

Architecture

This merged model is the base layer of a multi-tier inference stack:

[Query] -> Executive Controller (complexity routing)
              |
              v
[Merged Orchestrator Model]  <-- this repo
              |
              v
[LoRA Hot-Swap: newton, davinci, empathy, ...]
              |
              v
[Multi-Agent Debate + Semantic Tension]
              |
              v
[Coherence Check + AEGIS Ethics]
              |
              v
[Synthesized Response]

Related Repos

• Raiff1982/codette-llama-3.1-8b-gguf - Quantized GGUF for local inference
• Raiff1982/codette-lora-adapters - 9 LoRA perspective adapters
• Raiff1982/Codette-Reasoning - Training datasets

Training

The orchestrator adapter was trained with QLoRA:

• 4000 training examples covering routing, debate, coherence, synthesis
• 4 epochs on NVIDIA A10G (24GB)
• LoRA rank 16, alpha 32, dropout 0.05
• Merged using PeftModel.merge_and_unload()

License

Subject to the Llama 3.1 Community License.