finiarisab/tamgpt-orch

TamGPT Orchestrator

TamGPT Orchestrator — LoRA Training Repository TamGPT Orchestrator is a deterministic multimodal controller designed to route user requests to the correct tools, models, or subsystems. It is not a chatbot. It is trained to output strict JSON decisions that follow a predefined schema. This repository contains:

• The training script (train_orchestrator.py)
• The orchestration dataset (tamgpt_orchestrator_dataset.jsonl)
• The LoRA configuration
• Instructions for running training on Hugging Face GPU Training Jobs

🔧 Purpose TamGPT Orchestrator is built to:

• Analyze multimodal intent
• Select the correct tool or model
• Enforce deterministic routing
• Produce JSON‑only decisions
• Support commercial‑grade automation pipelines It is designed for systems where reliability, safety, and tool‑first reasoning matter more than open‑ended conversation.

📦 Repository Contents | | | | train_orchestrator.py | | | tamgpt_orchestrator_dataset.jsonl | | | requirements.txt | | | .gitattributes | |

🧠 Base Model Training is performed on: Qwen/Qwen2.5‑7B‑Instruct This model provides:

• Strong reasoning
• High‑quality instruction following
• Excellent JSON compliance
• Efficient LoRA fine‑tuning

🏋️ Training This repository is designed to run on Hugging Face Training Jobs using a GPU such as:

• A10G
• A100
• T4 (slower) Entry point train_orchestrator.py

Arguments --dataset_path tamgpt_orchestrator_dataset.jsonl --output_dir ./outputs

Dependencies Automatically installed from: requirements.txt

📤 Outputs Training produces a LoRA adapter containing:

• adapter_model.safetensors
• adapter_config.json
• Tokenizer files These can be downloaded from the Training Job artifacts and deployed in any inference environment.

🚀 Usage (Inference) To load the trained orchestrator: from transformers import AutoModelForCausalLM, AutoTokenizer from peft import PeftModel

base = "Qwen/Qwen2.5-7B-Instruct" adapter = "finiarisab/tamgpt-orchestrator"

tokenizer = AutoTokenizer.from_pretrained(base, trust_remote_code=True) model = AutoModelForCausalLM.from_pretrained(base, trust_remote_code=True) model = PeftModel.from_pretrained(model, adapter)

Then generate: output = model.generate( tokenizer(prompt, return_tensors="pt").input_ids, max_new_tokens=512 ) print(tokenizer.decode(output[0]))

📘 Dataset Format Each entry contains:

• Conversation history
• Multimodal intent analysis
• Capability routing context
• Available tools
• Telemetry
• Ground‑truth JSON decision The training script converts each entry into a strict: PROMPT → JSON decision

pair.

🔒 License & Commercial Use This repository is intended for private, commercial deployment. Model weights, dataset, and training outputs are restricted to authorized users.