VIC-2B

A small model that turns a question about a data file into a typed plan, which the vasuki engine checks and runs on your machine.

Ask in plain words, get the answer as a table. The model never does arithmetic, never writes code, and never sees the whole file: it is shown column names, types and a few sample rows, and it writes a short plan of steps such as filter, aggregate and join. The engine does the computing.

Use

VIC-2B is not a chat model. Use it through vasuki, which sends it the right input and runs the plan it writes. With uv installed, one line:

uvx --from git+https://github.com/rahul05ranjan/vasuki-vic vasuki ask "Total amount per vendor, highest first." -i invoices.csv --ollama --model rahuul/vic-2b

{
  "status": "ok",
  "columns": ["vendor", "total_amount"],
  "row_count": 3,
  "rows": [["Acme", 2500.0], ["Bolt", 500.5], ["Cora", 0]]
}

Add --show-plan to see the plan the model wrote. Inputs can be .csv, .tsv, .json or .jsonl; give two -i for a join.

Calling it yourself

Two things to know if you call the model without vasuki:

• Turn reasoning off. Ollama turns it on by default for this model family, and VIC asked to reason returns no plan. Send "think": false on Ollama’s API, or "reasoning_effort": "none" on the OpenAI-style endpoint. With ollama run, pass --think=false.
• The user message is JSON. The request plus a profile of each table, as vasuki.request.planner_input builds it. The system prompt is stored with this model.

The reply is one JSON object: a plan, a needs_clarification question, or an unsupported reason.

How good it is

Scored by running the model’s plan and comparing the resulting table with the expected one. The reference is gpt-oss-120b, a 120B-parameter model, given the full format description.

These figures were measured on the 6-bit MLX build. This 8-bit GGUF build scores the same as that build on the development set (210 of 245).

About one plain-language request in seven gets a plan that runs and returns the wrong table, with no warning. Read the plan when the answer matters.

Details