Pomoika24/C1-Tachu

Overview

C1-Tachu is a specialized 14.8 billion parameter coding AI, fine-tuned from Qwen2.5-Coder-14B by Pomoika24. Its core mission is to minimize the time from prompt to production-ready software by optimizing for developer throughput rather than benchmark scores. The model leverages a QLoRA training approach across eight stages, building skills in various software development domains.

Key Capabilities

• Fast Code Generation: Produces correct implementations with minimal, concise code.
• Debugging Speed: Quickly identifies root causes, trained on real GitHub commit-fix pairs.
• Project Navigation: Answers questions about unfamiliar repositories without extensive file reading.
• Rapid Architecture: Streamlines the planning-to-coding loop, from requirements to code skeleton.
• Self Verification: Utilizes an agent loop with tool use (e.g., run_tests, read_file) to achieve working solutions.
• Agent Workflows: Supports complex multi-step task decomposition and parallel tool calls.

Training Details

The model was trained using QLoRA (4-bit nf4 quantization) via the Axolotl framework. Training involved 8 stages, with each stage training a fresh LoRA adapter on the previous stage's merged model. Notably, Stage 7 (Preference Optimization) was skipped due to time constraints, meaning the model may sometimes produce more verbose solutions.

Security Posture

C1-Tachu maintains a passive security stance, meaning it defaults to safe coding patterns (e.g., parameterized queries, input validation) and avoids generating known-vulnerable code. It does not actively scan for vulnerabilities but will not exacerbate existing ones.

Intended Use

C1-Tachu is designed for software engineers, coding agents, and assistants to accelerate workflows, automate debugging, and aid in architecture planning. It is not intended for automated security auditing or deployment in safety-critical systems without human review.