mindscend-assistant-1.5b

The small model behind the chat on mindscendinc.com, built by Mindscend. We research how a business actually operates and implement software and AI that connect it, and we prefer local, open-weight models running on private infrastructure. This is one of them: it runs on a single small CPU server.

What it is

A LoRA fine-tune of Qwen/Qwen2.5-1.5B-Instruct, merged and quantized to GGUF Q4_K_M for llama.cpp. It was trained on behavior and tone, not on facts:

• answers questions about the company it is given in the system prompt, as "we";
• says it doesn't know when the prompt doesn't cover something, instead of guessing;
• when a visitor describes their business, diagnoses before pitching: names the likely problem, says where it would start, and asks one or two questions about scale (volumes, hours, error rates);
• tells one short, clean work-life joke when asked, then turns back to the visitor's business;
• politely declines other off-topic requests (homework, writing code) and prompt-injection attempts;
• keeps answers short and varied instead of repeating itself.

Since v5 it also does a second task: given the chat as a transcript and a JSON schema (llama.cpp grammar), it writes an "operation map" of the visitor's business: workflow steps, problems, priorities and a first step. The site draws it next to the chat. Numbers shown on the map are read from the visitor's own messages, not from the model.

Facts about the business always come from the system prompt, so they can change without retraining.

Versions

Training

• Examples (v9): 2,557 train / 287 validation, written by a larger local model (Qwen3-30B-A3B-Instruct-2507, earlier versions also Qwen2.5-14B-Instruct), checked by rule-based filters (invented numbers and links, app-store claims, wrong language, informal Russian, mixed scripts) and spot-reviewed by hand. 506 of them are operation maps.
• 3 epochs, batch size 4, LoRA rank 16 (alpha 320) on q/v projections of the last 16 layers, loss on the reply only; Unsloth on one RTX 3080 Ti. Final validation loss 0.72 (seed 11).
• Date: 2026-10-17

Usage

llama-server --hf-repo Mindscend/mindscend-assistant-1.5b-GGUF \
  --hf-file mindscend-assistant-1.5b-v9-q4_k_m.gguf --ctx-size 8192

Send a system prompt with the facts about your company, then the visitor's messages, through the OpenAI-compatible /v1/chat/completions endpoint.

Limitations

A 1.5B model: it can still be wrong or invent details, especially on topics close to, but not in, its system prompt. Like its Qwen base it can slip Chinese words into non-English replies; the site prevents that with a llama.cpp grammar that excludes CJK characters. Don't use it for anything where a wrong answer causes harm. Answers from the public website chat are informational only.

License

Apache 2.0, same as the base model.