vasuki: ask a table a question in plain words

VIC-2B-GGUF

VIC-2B as a GGUF file, for Ollama and llama.cpp on macOS, Linux and Windows.

VIC-2B turns a question about a data file into a typed plan, which the vasuki engine checks and runs on your machine. The full description, evaluation and limits are on the main model card.

File Quantisation Size
VIC-2B-Q8_0.gguf 8-bit (Q8_0) 2.0 GB

Use with vasuki

With uv and Ollama installed, one line. The model is fetched into Ollama the first time.

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

Use with Ollama directly

ollama pull hf.co/rahul05ranjan/VIC-2B-GGUF:Q8_0

Two things to know when calling it without vasuki:

  • Turn reasoning off. Ollama turns it on by default for this model family, and VIC asked to reason argues with the request and returns no plan. Send "think": false on Ollama's API, or "reasoning_effort": "none" on the OpenAI-style endpoint.
  • It is not a chat model. It expects vasuki's system prompt and, as the user message, the request with a profile of the tables as JSON. vasuki ask sends both; see the main model card.

Accuracy of this build

On vasuki's development set (evals/v2, 245 tasks), this file run through Ollama scores 210, the same as the 6-bit MLX build. The figures on the main model card were measured on the MLX build.

How it was made

The merged half-precision VIC-2B weights were converted with llama.cpp's convert_hf_to_gguf.py --outtype q8_0 --no-mtp. The base model's extra multi-token-prediction layer is not part of VIC-2B and is left out.

Licence

Apache License 2.0. See the main model card for what the model was trained on.

Downloads last month
30
GGUF
Model size
2B params
Architecture
qwen35
Hardware compatibility
Log In to add your hardware

8-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for rahul05ranjan/VIC-2B-GGUF

Finetuned
Qwen/Qwen3.5-2B
Quantized
(1)
this model