mtcicero26/fiberbrowser-copilot-1.5b-v1

Overview

mtcicero26/fiberbrowser-copilot-1.5b-v1 is a 1.5 billion parameter, LoRA-fine-tuned version of Qwen/Qwen2.5-Coder-1.5B-Instruct. It functions as a 'tiny' tier copilot for the FiberBrowser, a local genome-browser planner. The model's primary role is to convert natural language analysis requests into structured JSON action plans, which are then validated and executed by the browser.

Key Capabilities

• Genome-Browser Planning: Specializes in generating structured JSON action plans for genomic analysis within FiberBrowser.
• Action Vocabulary: Supports 11 distinct action types, including navigate_gene, scan_peaks, cluster_selected_peaks, add_region_label_rule, query_motif_overlaps, and an api_call escape hatch.
• JSON Output: Emits JSON action plans encapsulated within <think>...</think> tags, which are validated by the browser.
• Local Deployment: Designed for local execution, with recommendations for Macs with 16+ GB unified memory.

Training Details

The model was fine-tuned using LoRA distillation, with Claude (Anthropic) serving as the teacher to generate structured request-to-plan demonstrations. The training dataset comprised 319 hand-curated and LLM-distilled examples across 24 workflow categories, including navigation, peak detection, clustering, and motif overlaps. Training was performed on Apple Silicon Macs using the MLX framework.

Known Limitations

• Trained on synthetic and Claude-distilled data; real organic usage data is being collected.
• May drop tail actions in multi-step plans exceeding 5 actions.
• Agentic accuracy might be lower as LoRA was trained on single-shot plans, not tool-call traces.