Saif 4B v2
A fast decision model. Send it a state and a set of typed questions; it returns a probability for every answer in one forward pass. No text generation, no chat — it only makes decisions. Speaks the System One / Jev typed-decision format.
- Decision Index 0.3, public suite: 57.71 (raw 68.61), complete, no unsupported requests.
- Built on Qwen/Qwen3.5-4B, full fine-tune.
Results
Decision Index 0.3, public suite, run with the official kit.
| Area | Saif 4B v2 |
|---|---|
| Overall index | 57.71 |
| Tools & Automation | 83.1 |
| Retrieval & Classification | 59.9 |
| Language Understanding | 63.4 |
| Arts & Human Taste | 53.1 |
| Knowledge & Reasoning | 34.2 |
Strongest benchmarks: BFCL 96.9, FinEntity 92.7, HellaSwag 91.8, CLINC150+OOS 91.2, Home appliance simulator 86.4, API-Bank 86.2, BANKING77 83.9, BPoMP 82.3. Scores are chance-corrected skill (0 = random guessing, 100 = perfect).
Quick start
import sys
sys.path.insert(0, '/path/to/saif-4b-v2')
from decision_index_engine import SaifEngine # needs the decision-index kit on PYTHONPATH
engine = SaifEngine(model='/path/to/saif-4b-v2', device='cuda')
result, _ = engine('The parcel arrived broken and I want my money back today.', {
'refund': {'type': 'noul', 'instructions': 'Should the agent offer a refund?'},
'tone': {'type': 'choice', 'criteria': {'angry': 'Angry', 'neutral': 'Neutral', 'happy': 'Happy'}},
})
print(result['answers'])
Requirements: Python 3.12, torch 2.8, transformers with Qwen3.5 support, safetensors, accelerate.
License
Weights released under CC-BY-NC-4.0 (non-commercial), because some training sources are research-only. Base model: Qwen/Qwen3.5-4B, Apache-2.0.
- Downloads last month
- 150