RichWoollcott/architect-agent-gemma4-26b-moe

Architect Agent (Gemma-4-26B-A4B MoE)

A fine-tune of unsloth/gemma-4-26b-a4b-it that emulates a senior software architect: applies DDD strategic patterns, surfaces architectural tensions, and reasons explicitly via blocks before recommending.

Fine-tuned from unsloth/gemma-4-26b-a4b-it using Unsloth + TRL on a single NVIDIA DGX Spark GB10 (121 GB unified memory).

Intended use

Pairing with a developer on architecture decisions: bounded-context design, integration patterns, sequencing legacy-system migrations, and articulating trade-offs in writing. Built to be served behind llama-swap / Ollama / vLLM as part of a multi-agent system.

Sample prompt

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

repo = "RichWoollcott/architect-agent-gemma4-26b-moe"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(
    repo, torch_dtype=torch.bfloat16, device_map="cuda"
)

messages = [{"role": "user", "content": "I'm building a new payments service that needs to integrate with five legacy systems. Which DDD strategic patterns should I apply first and why?"}]
inputs = tok.apply_chat_template(
    messages, return_tensors="pt", add_generation_prompt=True
).to("cuda")
out = model.generate(inputs, max_new_tokens=600, do_sample=False)
print(tok.decode(out[0][inputs.shape[1]:], skip_special_tokens=False))

The model is trained to open every response with a <think>...</think> reasoning block followed by the visible answer. If you do not see the tags, add Begin every response with <think>...</think> to the system prompt.

Limitations

Trained on ~900 synthetic dialogues generated by an adversarial Player-Coach pipeline; coverage is biased towards the 10 architect-dimension categories enumerated in the dataset-factory generation plan. The model has no first-hand engineering experience and will confidently reason from incorrect premises if those premises are accepted. Always pair with a RAG layer over authoritative architecture sources for factual grounding.

Training data

894 reasoning-style (<think>...</think> then answer) dialogues covering 10 DDD-flavoured architect dimensions, accepted by an LLM Coach with a structured-JSON rubric. Generated by the agentic-dataset-factory Player-Coach pipeline. 83.2% acceptance rate; zero chat-template-token leaks; 100% <think> coverage.

The data is not released alongside the weights pending a license review of source material referenced during generation. The generator (agentic-dataset-factory, Player-Coach adversarial pipeline) is open-source and reproducible.

Training procedure

• Framework: Unsloth + TRL (SFT)
• Precision: bfloat16
• PEFT: LoRA, rank 16
• Optimiser: AdamW (Unsloth defaults)
• Epochs: 1, effective batch size 4, max-seq-length 2048
• Hardware: NVIDIA DGX Spark GB10 (single device, unified memory)
• Container: nvcr.io/nvidia/pytorch:25.11-py3
• Verified-compatible pins: transformers==5.5.4, accelerate==1.10.0, trl==0.26.1, datasets==4.3.0, latest unsloth/unsloth_zoo/bitsandbytes.

Loss trajectory and reproduction recipe: see the source runbook in appmilla/agentic-dataset-factory — domains/architect-agent/RUNBOOK-architect-fine-tune.md.

Artefacts in this repo

• / (root) — merged-16bit weights (full standalone model, ~49 GB). Load with AutoModelForCausalLM.from_pretrained(repo_id).
• lora-adapter/ — the LoRA adapter (~1.9 GB). Apply on top of the base model with PeftModel.from_pretrained(base, repo_id, subfolder="lora-adapter").

GGUF quantisations (q4_k_m, BF16) live in a separate repo: RichWoollcott/architect-agent-gemma4-26b-moe-GGUF.

License

Released under the Gemma Terms of Use, as required by the base model.

Citation

If you use this model, please cite the base model, TRL, and Unsloth:

@misc{vonwerra2022trl,
  title  = {TRL: Transformer Reinforcement Learning},
  author = {Leandro von Werra and others},
  year   = 2020,
  url    = {https://github.com/huggingface/trl}
}