saif-4b / README.md
GyrTech's picture
Saif 4B v2.1 (Decision Index 0.3 public 62.94)
8319c9c verified
|
Raw History Blame Contribute Delete
3.09 kB
---
license: cc-by-nc-4.0
base_model: Qwen/Qwen3.5-4B
library_name: transformers
pipeline_tag: text-classification
tags: [typed-decisions, decision-model, system-one, jev-compatible, custom-code]
---
# Saif 4B (v2.1)
**A fast decision model.** Send it a state and a set of typed questions in one request; it returns a probability for
every answer in a single forward pass. No text generation and no chat: it only makes decisions. It speaks the System
One / Jev typed-decision format.
- **Decision Index 0.3, public suite: 62.94** (raw 72.21), all 140,178 requests answered, none unsupported.
v2.0 scored 57.71.
- **Fast:** 32 ms median, 107 ms mean, 64 ms p80 per request,
one request at a time.
- **Built on Qwen/Qwen3.5-4B**, full fine-tune, block-bidirectional attention in the full-attention layers.
Earlier versions stay available: `v1.0` and `v2.0` are tagged in this repository.
## Results
Decision Index 0.3, public suite (37 benchmarks), run with the official kit on one RTX PRO 6000, one pass per request.
Scores are chance-corrected skill (0 is random guessing, 100 is perfect).
| Area | v2.1 |
| --- | ---: |
| **Overall index** | **62.94** |
| Tools & Automation | 83.7 |
| Retrieval & Classification | 68.6 |
| Language Understanding | 68.1 |
| Arts & Human Taste | 56.0 |
| Knowledge & Reasoning | 41.4 |
Strongest benchmarks: BFCL 97.2, CLINC150+OOS 92.7, HellaSwag 92.3, FinEntity 91.3, BANKING77 90.8, API-Bank 87.2, BPoMP 85.2, When2Call MCQ 82.7, Home appliance simulator 81.8, POP909-CL 80.6.
Weakest: HLE, ChessBench, GPQA Diamond, Humicroedit, SATA-Bench, BRIGHT.
## Quick start
```python
import sys
sys.path.insert(0, '/path/to/saif-4b')
from decision_index_engine import SaifEngine # needs the decision-index kit on PYTHONPATH
engine = SaifEngine(model='/path/to/saif-4b', 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.
## Training
Full fine-tuning of Qwen/Qwen3.5-4B (revision `851bf6e8`) in rounds at a constant learning rate on typed decisions
drawn only from public training splits plus programmatically generated decisions, with audited, steered data mixes;
then a 1,000-step cosine cooldown. Output temperatures and a PhishNChips decision boundary are fitted on held-out
splits of the training data. Labels are the datasets' own human/source annotations or computed by code. Every training
directory is screened against all Decision Index suite items, the evaluation panel and calibration dev sets before use;
no Decision Index item is used for training or selection.
## License
Weights released under CC-BY-NC-4.0 (non-commercial), because some training sources are research-only. Base model:
[Qwen/Qwen3.5-4B](https://huggingface.co/Qwen/Qwen3.5-4B), Apache-2.0.