Text Classification
Transformers
Safetensors
English
Chinese
lemondecider_decision
feature-extraction
lemondecider
decision-engine
typed-decision
classification
custom_code
Instructions to use inclusionAI/LemonDecider with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use inclusionAI/LemonDecider with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-classification", model="inclusionAI/LemonDecider", trust_remote_code=True)# pip install -U transformers accelerate # Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("inclusionAI/LemonDecider", trust_remote_code=True, device_map="auto") - Notebooks
- Google Colab
- Kaggle
Update model card (2026-11-11)
Browse files
README.md
CHANGED
|
@@ -13,7 +13,7 @@ tags:
|
|
| 13 |
|
| 14 |
# LemonDecider 4B
|
| 15 |
|
| 16 |
-
**Model ID:** `lemondecider-4b-
|
| 17 |
**Base model:** `Qwen/Qwen3.5-4B`
|
| 18 |
|
| 19 |
This v3 release makes typed decisions from a shared text state and named questions. It returns choice probabilities, yes/no probabilities (noul), or ordinal scores without generating an answer token stream. Requests that fit use one joint forward; overlength question lists are recursively bisected, preserving the complete state and every option in each chunk. Inputs are never truncated.
|
|
@@ -24,7 +24,7 @@ This v3 release makes typed decisions from a shared text state and named questio
|
|
| 24 |
|---|---|
|
| 25 |
| Architecture | Qwen3.5-4B hybrid backbone with a 255-option decision readout |
|
| 26 |
| Loader | `modeling_lemondecider.LemonDeciderForDecisionClassification`, `trust_remote_code=True` |
|
| 27 |
-
| Model type / protocol | `lemondecider_decision` / `
|
| 28 |
| Attention | Noncausal periodic full-attention layers; hybrid recurrent layers; SDPA |
|
| 29 |
| Readout | Native inline joint-question readout (`joint_readout.predict_joint`) |
|
| 30 |
| Temperature | 1.0026751894886363, inherited from the parent checkpoint; not refit |
|
|
@@ -82,10 +82,10 @@ recommended reproducing CLI.
|
|
| 82 |
|
| 83 |
## Evaluation results
|
| 84 |
|
| 85 |
-
### 1) Decision Index, edition 0.
|
| 86 |
|
| 87 |
On **2026-11-04**, this model was evaluated over the complete public
|
| 88 |
-
0.
|
| 89 |
iterative question bisection. The official scorer reported `complete: true`.
|
| 90 |
The underlying selected file contains **140,620 rows**: **140,178 scoreable**
|
| 91 |
requests and **442 officially excluded** requests. The benchmark-native summary
|
|
@@ -94,7 +94,7 @@ remain in index coverage.
|
|
| 94 |
|
| 95 |
| Headline | Value |
|
| 96 |
|---|---:|
|
| 97 |
-
| **Decision Index (public 0.
|
| 98 |
| Raw index | 70.84 |
|
| 99 |
| Breadth skill | 60.28 |
|
| 100 |
| Median / mean / p95 request latency | 49.9 ms / 121.2 ms / 286.9 ms |
|
|
|
|
| 13 |
|
| 14 |
# LemonDecider 4B
|
| 15 |
|
| 16 |
+
**Model ID:** `lemondecider-4b-20261104-v3`
|
| 17 |
**Base model:** `Qwen/Qwen3.5-4B`
|
| 18 |
|
| 19 |
This v3 release makes typed decisions from a shared text state and named questions. It returns choice probabilities, yes/no probabilities (noul), or ordinal scores without generating an answer token stream. Requests that fit use one joint forward; overlength question lists are recursively bisected, preserving the complete state and every option in each chunk. Inputs are never truncated.
|
|
|
|
| 24 |
|---|---|
|
| 25 |
| Architecture | Qwen3.5-4B hybrid backbone with a 255-option decision readout |
|
| 26 |
| Loader | `modeling_lemondecider.LemonDeciderForDecisionClassification`, `trust_remote_code=True` |
|
| 27 |
+
| Model type / protocol | `lemondecider_decision` / `lemondecider-v1` |
|
| 28 |
| Attention | Noncausal periodic full-attention layers; hybrid recurrent layers; SDPA |
|
| 29 |
| Readout | Native inline joint-question readout (`joint_readout.predict_joint`) |
|
| 30 |
| Temperature | 1.0026751894886363, inherited from the parent checkpoint; not refit |
|
|
|
|
| 82 |
|
| 83 |
## Evaluation results
|
| 84 |
|
| 85 |
+
### 1) Decision Index, edition 0.4 — full public suite
|
| 86 |
|
| 87 |
On **2026-11-04**, this model was evaluated over the complete public
|
| 88 |
+
0.4 suite using **16 PPU-ZW810E workers**. The run was joint inference plus
|
| 89 |
iterative question bisection. The official scorer reported `complete: true`.
|
| 90 |
The underlying selected file contains **140,620 rows**: **140,178 scoreable**
|
| 91 |
requests and **442 officially excluded** requests. The benchmark-native summary
|
|
|
|
| 94 |
|
| 95 |
| Headline | Value |
|
| 96 |
|---|---:|
|
| 97 |
+
| **Decision Index (public 0.4)** | **61.67** |
|
| 98 |
| Raw index | 70.84 |
|
| 99 |
| Breadth skill | 60.28 |
|
| 100 |
| Median / mean / p95 request latency | 49.9 ms / 121.2 ms / 286.9 ms |
|