File size: 4,276 Bytes
3546429 fd91598 3546429 fd91598 3546429 fd91598 3546429 fd91598 0b67e32 fd91598 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 | ---
title: IDU
emoji: 🧠
colorFrom: purple
colorTo: gray
sdk: gradio
sdk_version: 5.49.1
app_file: app.py
pinned: false
license: apache-2.0
short_description: Typed decisions (choice/score/noul) on open weights
---
# IDU
**JEV** shifted one letter back — the same shift that turns **IBM** into **HAL**.
A typed-decision service in Jev's shape: give it a `state` and typed `questions`, get typed
answers with probabilities and confidence. No text generation, so nothing to parse and
nothing to hallucinate.
| type | question | answer |
|---|---|---|
| **choice** | which of these options? | the option, a probability per option, confidence |
| **score** | where on this rubric? | a position along your levels (can land between them), distribution |
| **noul** | is this true? | one number: the probability that it is |
## Backend
[`convaiinnovations/laya`](https://huggingface.co/convaiinnovations/laya) — 421M
ModernBERT-large with a decision head, trained by RL against strictly proper scoring rules
(RLCD). Apache 2.0.
Why this one:
- **It already is the Jev shape.** `state` + typed questions in, typed decisions out. No
adapter, no tool-calling workaround, no trigger vocabulary.
- **Options are judged independently.** Each option is scored at its own `[MASK]` token and
then softmaxed over that question's options. That is what Jev documents ("each level is
judged on its own against the state"), and it removes the polarity problem a tool-calling
model has: a `noul` question is two markers, not two competing armed tools.
- **One forward pass** answers every question in a call.
- **PyTorch**, so it fits ZeroGPU natively — none of the JAX dtype/preallocation work it took
to get a JAX model running on this runtime.
## Endpoints
- `POST /systemone` (Gradio API name `systemone`) — the Jev shape.
- `/triage` — four-question support triage in one call, with a confidence-gated action line.
- `/compare` — head-to-head against the tuned Needle 3 adapter on `dkappe/needle3-gpu`.
- `/health` — device, temperature in force, warm-up result.
## Honest status
**The base checkpoints are near chance zero-shot** on their own typed-decisions benchmark:
0.362 against a 0.318 random and a **0.461 majority-class baseline**. The card's headline
0.766 belongs to a checkpoint fine-tuned on that benchmark's own training split, and it says
so itself: *"Laya is a fast base to specialise, not a zero-shot decision engine."*
**The probabilities ship already calibrated.** The model card reports mean ECE 0.466 falling
to 0.081 after fitting one temperature per (question type, option count) — and that fitting
is baked into the checkpoint: `rl_agent_config.json` carries `temperature [1.6369, 1.2514,
1.9834]` for choice/score/noul plus per-option-count buckets. So the defaults here *are* the
calibrated path. The temperature controls use **0 = keep the checkpoint's value** as their
sentinel; setting a positive number overrides deliberately.
**`score` is the weakest primitive** by the card's own numbers (SST-5 0.372).
**High-cardinality `choice` degrades** because options share a fixed token budget: 0.425 on
Banking77 (77 options) vs Jev's 0.870. Fine at our 3–4 options; raise `head_max_len` or use a
coarse-to-fine split well before 50.
**English only** at the repo root. `laya-multilingual` (mmBERT-base) covers 100+ languages
and is ~2.2x faster.
## Implementation notes worth keeping
- **`spaces` must be imported before `torch`.** The spaces library patches torch's CUDA
initialisation so a GPU can be attached per call; import torch first and every decorated
call dies inside `worker_init` with `RuntimeError: No CUDA GPUs are available`. Nothing
that pulls in torch — gradio, transformers, laya, safetensors — may appear above that
guard block. Taken from the model's own demo Space, which documents the same trap.
- **Assets load at module scope**, before the UI is reachable, because network access inside
a `@spaces.GPU` function is restricted. Warm-up runs at startup so the first click is not
paying model construction.
- **One decorated entry point** for every tab, so the start-up warm-up covers all of them
instead of each tab paying its own GPU attach.
Apache 2.0 — weights and SDK both.
|