faysalbenahmed/AMF-G1-4B-Experimental
AMF-G1-4B-Experimental is a 4 billion parameter language model developed by Fayçal Benahmed of Stack Moderne, derived from Qwen3-4B-Instruct-2507 with supervised LoRA training. It features a 32768 token context length. This model serves as an experimental artifact demonstrating the AI Mission Foundry (AMF) fabrication process, specifically its ability to produce learned neural material, rather than being a production-qualified model. Its primary purpose is to provide evidence of an incomplete optimization loop and measured failure as part of research into autonomous AI system development.
Loading preview...
AMF G1 — 4B Experimental Overview
AMF-G1-4B-Experimental is a 4 billion parameter language model created by Fayçal Benahmed of Stack Moderne, serving as the first learned neural artifact fabricated by the AI Mission Foundry (AMF). Derived from Qwen/Qwen3-4B-Instruct-2507, it underwent supervised LoRA training with specific parameters (rank=16, learning rate=4e-5, 72 steps, seed=190901) followed by a safe merge. This model is explicitly published as evidence of the AMF fabrication process, not as a qualified production model, and demonstrates the ability to produce learned neural material.
Key Characteristics & Experimental Context
- Experimental Artifact: G1 is a research artifact, showcasing the AMF's ability to fabricate neural models, rather than a fully optimized or production-ready system. It was produced before the AMF's autonomous optimization and REPAIR loop was fully implemented.
- Measured Failure: The model's qualification result was
FAIL_OBSERVEDwith aqualified_outputof 0.3472, which is deliberately preserved as part of the research record to inform future development. - Local Deployment Viability: Despite its experimental status, the canonical BF16 model successfully executed locally on CPU, and a GGUF Q5_K_M derivative was created, demonstrating local runtime viability with measured throughputs (11.73 tokens/s prompt, 5.68 tokens/s generation on AMD Ryzen 5 8540U).
- Behavioral Gate Pass: A separate local deployment-oriented micro-evaluation, testing aspects like
SUPPORTED_FACTSandEXACT_EVIDENCE_FIDELITY, resulted in aPASS(15/15 exact passes), though this does not alter its overall experimental qualification status.
Good for
- Understanding AI Fabrication Processes: Ideal for researchers and developers interested in the methodology of AI model creation, particularly the AMF's approach to fabricating learned neural material and iterative optimization.
- Studying Model Limitations and Failures: Useful for analyzing how early-stage experimental models perform and the insights gained from their measured failures in a research context.
- Exploring Local Inference: Provides a case study for local CPU-only inference with a 4B parameter model, including GGUF derivative performance metrics.