Ryannnnnnz/SuperNova-Tuned

SuperNova Tuned

Tuned multimodal checkpoint.

Model

• Base: sensenova/SenseNova-U1-8B-MoT-Interleaved
• Training: full-parameter SFT, not LoRA/PEFT
• Checkpoint: step 163, after 3 epochs over the mixed fake-search corpus
• Training examples: 4,546 image-search and 1,739 text-search examples
• Conditioning: search input only; no question or ground-truth answer is provided to the simulator

Behavior contracts

Text search

Input:

[Text Search Query] QUERY

Expected output: an [Text Search Results] block containing up to five title, snippet, and URL entries.

Image search

Input: one query image and the prompt <image>.

Expected output: an [Image Search Results] block containing exactly five interleaved title and generated-image entries.

Inference

Use the SenseNova-U1 code and its interleaved inference entrypoint:

git clone https://github.com/OpenSenseNova/SenseNova-U1.git
cd SenseNova-U1
pip install -e .

python examples/interleave/inference.py \
  --model_path Ryannnnnnz/SuperNova-Tuned \
  --prompt "[Text Search Query] Acorn Weevil identification" \
  --system_message "You are a Google Search (Serper API) simulator. Based only on the user's search query, generate up to 5 plausible relevant documents. Output ONLY the [Text Search Results] block." \
  --no-think_mode

For image search, pass --prompt '<image>' and --image /path/to/query.jpg with the image-search system message.

Status

This checkpoint is intended for research and format/protocol experiments. It learns both text-search and image-search output modes. Semantic relevance, output diversity, and strict parser compatibility have not been established at production quality.