fiberbrowser-copilot-3b-v1-q4

Copilot 3B q4 (light, quantized)

LoRA-fine-tuned (and int4 quantized) Qwen/Qwen2.5-Coder-3B-Instruct for the FiberBrowser Copilot — a local genome-browser planner that turns natural-language analysis requests into structured JSON action plans.

Tier: light (3.0B params, int4 quantized). Recommended for Macs with 12+ GB unified memory (6 GB minimum).

Use

This is a Copilot tier — a single mlx_lm.server instance loads it and the FiberBrowser frontend talks to it via OpenAI-compatible chat completions. The browser handles the system prompt, viewport context, dataset metadata, and previous-turn history; the model emits a JSON action plan inside <think>...</think> tags.

# Install
pip install mlx-lm fiberbrowser   # or your preferred install

# Serve
python -m mlx_lm server --model mtcicero26/fiberbrowser-copilot-3b-v1-q4 --port 8080

# Point the browser at it
FIBERBROWSER_COPILOT_BACKEND=mlx \
FIBERBROWSER_COPILOT_MLX_URL=http://127.0.0.1:8080/v1/chat/completions \
python browser.py

Or simply: python browser.py --copilot mlx — the launcher auto-detects RAM, downloads the right tier, starts the server, and shuts it down with the browser.

Training

Supervised fine-tuning via LoRA distillation:

• Teacher: Claude (Anthropic) generated structured (request → plan) demonstrations.
• Dataset: 319 hand-curated and LLM-distilled examples across 24 workflow categories — navigation, peak detection, clustering, footprint analysis, motif overlaps, primer design, contact maps, annotations, capture regions, multi-step composites, and edge cases.
• Method: LoRA, rank 16, top 16 layers, 900 iters, lr 1e-4, batch size 1, grad checkpointing. Final val loss 0.016.
• Hardware: Apple Silicon Mac (MLX framework).

Output protocol:

<think>
{reasoning about which actions to chain}
</think>
{"summary": "...", "actions": [{"type": "...", ...}], "warnings": []}

The browser validates every plan through the same Python validator that produced the training data — any action with malformed coordinates, unsafe API paths, or unknown gene names is dropped before execution.

Action vocabulary

11 action types in the trained vocabulary:

• navigate_gene, navigate_region — viewport
• scan_peaks, set_peaks_locked, select_peak_at — peaks
• cluster_selected_peaks, cluster_peak_at_gene_tss — clustering
• add_region_label_rule — colormap rules (TSS-relative form for strand-aware coords)
• query_motif_overlaps — motif intersections
• api_call — escape hatch to any safe /api/... route (settings, render, sort, filter, footprint analyze, primer design, etc.)
• respond_to_user — terminator for the agentic loop

Known limitations

• Trained on synthetic + Claude-distilled demonstrations; real organic-usage data is being collected via opt-in session logging in the browser.
• Multi-step plans of >5 actions sometimes drop tail actions; can be partially mitigated with multi-turn conversation context.
• Tool-call (agentic) form is supported by the underlying base model but the LoRA was trained on single-shot plans, not tool-call traces — agentic accuracy may be lower.

Other tiers

Citation

If this model is useful in your work, please cite the FiberBrowser repository.

License

Apache 2.0 (matching the Qwen2.5-Coder base model license).