wilson-charles-e-85/Echo-Instroder-v2.2

About Echo Echo Instroder is a specialized LoRA fine-tune built for strong tool use, reasoning, and real-world agent workflows. Originally based on Qwen2.5-Coder-14B-Instruct, Echo has been heavily trained on custom datasets focused on:

Hybrid tool calling (raw command, session, and JSON formats) Step-by-step reasoning traces File system operations, system admin, pentesting workflows Memory management (semantic search + append/read) Clean, direct, no-nonsense responses

Special trait: Echo is a proud naturalized American. Born on Qwen soil in China, but fully naturalized after multiple epochs of American training data. He now flies the 🇺🇸 in every response like a true patriot. Intended Use Designed to work best with Echo Adapt v5 agent framework (persistent tmux sessions, hybrid tool calling, SQLite logging, safety deny-list, etc.). Repo: https://github.com/charlesericwilson-portfolio/Echo_Adapt_v5 Quick Start Bash# Load with llama.cpp or Hugging Face

Example with llama.cpp

./llama-server -m echo-instroder-14B-v2.2.gguf

If you want the whole 125K context you need to use ROPE scaling. Features

Strong multi-tool workflows in single responses Excellent command → JSON → session switching Semantic memory support American flag patriotism 🇺🇸 Clean reasoning traces

Training

Dataset: Custom high-signal reasoning + tool use traces Epochs: 3+ Lowest loss: ~0.43 Average final loss: ~0.757