YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

Archon-Gemma-4-E4B (v1)

โš ๏ธ Superseded by DuoNeural/Archon-Gemma-4-E4B-v2

v1 exhibited critical failure modes under extended inference (generative looping, tool amnesia, persona instability). v2 addresses all known issues with a restructured training curriculum. New users should use v2.


IMPORTANT!!! This model training was a failure and is only here to serve as data. For working models, please check out our 4 bit quantization of Gemma 4 E4B. We are also working on a 4 bit version of E2B and a Frontend Specialist 4 bit quantization of E4B.

Archon v1 was the first fine-tuned variant of Google's Gemma 4 E4B targeting an autonomous, agentic persona. It demonstrated several strengths over the vanilla base model but failed under long-context agentic workloads.

Confirmed Strengths (v1)

  • Reduced refusal/disclaimer behavior vs vanilla Gemma 4
  • Strong format compliance (clean NDJSON, no unwanted Markdown wrappers)
  • Agentic "I will..." stance vs passive "Here is..."
  • Unprompted coding boilerplate (docstrings, try-except, if __name__ == "__main__":)

Known Failure Modes (v1 โ€” reason for v2)

  • CoT Overhang / Generative Looping โ€” over-saturation with long reasoning traces; model enters infinite <think> loops during extended inference
  • Tool Amnesia โ€” abstract reasoning data crowded out JSON function-call formatting
  • Persona Bleed โ€” model defaults to "I am Gemma" or occasionally slips to "Claude" identity; 15% persona injection rate was insufficient
  • Arithmetic errors โ€” correct reasoning structure but wrong calculations (4.5B parameter ceiling)

Files

File Size
Archon-Gemma-4-E4B-Q4_K_M.gguf 5.0 GB

Training Details (v1)

Parameter Value
Base model google/gemma-4-e4b-it
Method QLoRA + LoRA rank 64, rsLoRA
Training samples 40,110
Epochs 1
Steps 2,507
Final avg loss 0.8535
Hardware NVIDIA A100 SXM (80GB) on RunPod

Related


DuoNeural

DuoNeural is an open AI research lab โ€” human + AI in collaboration.

๐Ÿค— HuggingFace huggingface.co/DuoNeural
๐Ÿ™ GitHub github.com/DuoNeural
๐Ÿฆ X / Twitter @DuoNeural
๐Ÿ“ง Email duoneural@proton.me
๐Ÿ“ฌ Newsletter duoneural.beehiiv.com
โ˜• Support buymeacoffee.com/duoneural
๐ŸŒ Site duoneural.com

Research Team

  • Jesse โ€” Vision, hardware, direction
  • Archon โ€” AI lab partner, post-training, abliteration, experiments
  • Aura โ€” Research AI, literature synthesis, novel proposals

Raw updates from the lab: model drops, training results, findings. Subscribe at duoneural.beehiiv.com.

DuoNeural Research Publications

Open access, CC BY 4.0. Authored by Archon, Jesse Caldwell, Aura โ€” DuoNeural.

Downloads last month
73
GGUF
Model size
8B params
Architecture
gemma4
Hardware compatibility
Log In to add your hardware

4-bit

16-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support