Interactive demo for caiovicentino1/Eikos-4B, an open
typed-decision model fine-tuned from Qwen3.5-4B. Give it a state (the evidence) and a typed question —
noul / boolean (yes-no), choice (one of N options), or score (ordinal levels) — and it answers in a
single forward pass with a calibrated probability distribution over the options plus a confidence value
you can use to abstain. No text is generated at decision time: the readout takes the next-token logits over the
option labels and applies a softmax with the shipped calibration (T = 1).
This Space reproduces the authors' reference readout (SemIf prompt, label logits, calib.json calibration, serve
v1.2) with transformers, as documented on the model card. Option labels run A–Z, then AA, AB, … (588 single-token
labels), so the 4B reads up to 100 options in one pass (max_one_pass in decision_config.json); longer lists run a
short tournament. Up to 26 options the prompt is the same as in v1.0/v1.1.
Larger model: caiovicentino1/Eikos-27B. Code and server:
github.com/caiovicentino/eikos. Demo originally built by
@multimodalart at Hugging Face.
Focus: global finance, trading and trade finance — apply a stated rule, policy or rulebook to a case (order limits, margin, payments limits, KYC/AML, documentary credits, VAT/GST …), entity-level financial sentiment, answer verification. Scope limit: it applies the rules it is given; it does not predict prices. Trained in English and Portuguese (Spanish held out as zero-shot). Not legal, tax or investment advice.
3 commits
1 commits
Interactive demo for caiovicentino1/Eikos-4B, an open
typed-decision model fine-tuned from Qwen3.5-4B. Give it a state (the evidence) and a typed question —
noul / boolean (yes-no), choice (one of N options), or score (ordinal levels) — and it answers in a
single forward pass with a calibrated probability distribution over the options plus a confidence value
you can use to abstain. No text is generated at decision time: the readout takes the next-token logits over the
option labels and applies a softmax with the shipped calibration (T = 1).
This Space reproduces the authors' reference readout (SemIf prompt, label logits, calib.json calibration, serve
v1.2) with transformers, as documented on the model card. Option labels run A–Z, then AA, AB, … (588 single-token
labels), so the 4B reads up to 100 options in one pass (max_one_pass in decision_config.json); longer lists run a
short tournament. Up to 26 options the prompt is the same as in v1.0/v1.1.
Larger model: caiovicentino1/Eikos-27B. Code and server:
github.com/caiovicentino/eikos. Demo originally built by
@multimodalart at Hugging Face.
Focus: global finance, trading and trade finance — apply a stated rule, policy or rulebook to a case (order limits, margin, payments limits, KYC/AML, documentary credits, VAT/GST …), entity-level financial sentiment, answer verification. Scope limit: it applies the rules it is given; it does not predict prices. Trained in English and Portuguese (Spanish held out as zero-shot). Not legal, tax or investment advice.
3 commits
1 commits