cpu-training

MasterAI 1.0 - 1v1 Limit Texas Hold'em AI /德州AI / 一对一限注德州扑克AI / 一對多限注德州撲克AI - CFR solver, Monte Carlo, real-time decision engine

RL GRPO Finetuning on CPU

Out-of-core LoRA fine-tuning and expert-routing measurement for Kimi K3 (2.78T MoE) on a 7.6 GB laptop: 93 layers streamed from a USB disk, forward checked against an independent C implementation, gradients checked by finite differences

QSELM: 34.1M CPU LM, 215,771 training tok/s (8,529x measured Qwen training); sealed QA 90.6% vs Qwen3.5-0.8B 45.8%, memory 69.6% vs Qwen3-0.6B-FC 3.2%. No GPU. | QSELM：3408.7万参数CPU模型，训练21.6万token/s（Qwen训练实测的8,529倍）；封存问答90.6%对Qwen3.5-0.8B 45.8%，跨轮记忆69.6%对Qwen3-0.6B-FC 3.2%；无需GPU。

CPU-trained reasoning model pipeline. LoRA SFT + DPO on SmolLM2-360M, GSM8K math reasoning, single-laptop deployment.

ARION ALPHA 1 — bilingual (Persian/English) web-development LLM · Qwen3-0.6B + LoRA SFT · GGUF for Ollama/llama.cpp · مدل زبانی دوزبانه‌ی فارسی-انگلیسی برای توسعه‌ی وب

Local-first AI workspace with NP-DNA — a NeuroPlastic DNA Network for CPU-native training, memory, automation, and dashboard.

A 0.5B model fine-tuned on a laptop CPU that beats a frontier model at translating questions into a deterministic query DSL. Wrong answers fail to compile instead of becoming plausible numbers.

An end-to-end pipeline for training and deploying a lightweight math reasoning language model (Qwen2.5-0.5B). Features CPU-compatible Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and an interactive web interface built with Flask and Streamlit.

Teaching a machine to learn code the way humans do — curriculum-based AI training on a laptop, no GPU, no cloud, no massive datasets. Proving that thoughtful data design beats brute-force scale.

考勤规则指令微调（LoRA / Qwen2.5-0.5B）：规则引擎当 teacher 自动生成指令数据，CPU 无 GPU 环境跑通全链路。status 判定正确率 0% → 58.8%（17 条留出集）。

Interactive Training Dashboard & CAGS-Operator Verification for JamOne Nano.

Native ARM64 PyTorch on Windows on ARM (Snapdragon X Elite): the official wheel benchmarked and found slower than x86 emulation, plus the WSL2 path that actually trains. One-command emulation diagnostic included.

Fine-tunes Gemma-3-1b Math SFT on a custom JSONL dataset in <20min on CPU. Computes before/after eval.

A CPU-first neuro-symbolic language model framework with autonomous web data collection, self-improvement loops, and hardware auto-fitting.