Pivot

Pivot is a decision model developed by Q1z. It scores a supplied set of options and returns a choice with probabilities. It does not generate chat text.

Built on LiquidAI/LFM2.5-Encoder-350M with a trained DSBT option-set scorer. Context and options are encoded separately, mean-pooled, scored together, and normalized over the valid options.

Model details

• Parameters: 357,631,745
• Weight types: float32
• Format: Safetensors, full model including the decision scorer
• Outputs: choice, noul (Boolean choice), and discrete score distributions
• Base revision: b886781f7c6f10ca9b7096e21b83e30a073c2f39
• Saved checkpoint: epoch index 0, global step 3102
• Export verification: exact weight round-trip and typed output agreement PASS

Quick start

Install torch, transformers, safetensors, and numpy using the versions recorded in requirements.txt. Review the bundled custom code before trusting it.

from transformers import AutoModel, AutoTokenizer

repo = "Q1z/Pivot"
tokenizer = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModel.from_pretrained(repo, trust_remote_code=True).eval()
result = model.choose(tokenizer,
    context="My invoice has the wrong total.",
    options=["billing", "technical support", "sales"])
print(result)

For typed questions use model.decide(tokenizer, state=..., questions=...); see serving/example_request.json and predict_example.py. CPU is supported. For CUDA, call model.to("cuda") before inference. After downloading the complete repository, use its local path with local_files_only=True; no base-model weights are fetched during loading. Private repository access requires an authorized Hugging Face token.

Evaluation and limitations

JevBench results will be added after testing. This export runs no full benchmark. The smoke test verifies loading and output structure, not accuracy or calibration. Probabilities have not been independently validated for calibration. Multilingual quality and speed are not yet benchmarked. Supplied options determine the available answers; this is not an open-ended text generator.

Package

model.safetensors is the only weight file. config.json, tokenizer files, and bundled Python modules implement loading and inference. dsbt_config.yaml records the saved training configuration. provenance.json records the checkpoint hash. Training backups and optimizer state are not included. Keep the source best.pt separately if further training is needed. ONNX is not exported.

Project: https://trypivot.me