Sieve-31B
Model card for Sieve-31B, the accuracy-tier decision model. It returns probabilities for typed questions and does not generate text.
Hugging Face: sthanika-ai/Sieve-31B (Text Classification, Apache-2.0 for the adapter and code)
API model name: sieve-31b (versioned ID sieve-31b-v1), also used as the escalation model in sieve-auto.
Sieve-31B answers typed questions about a text or JSON state. It returns a probability for each option. It never generates text.
Key facts
| Item | Value |
|---|---|
| Base model | google/gemma-4-31B-it, revision 842da37 |
| Base model licence | Gemma terms, including Google's prohibited-use policy. These flow down to customers (see Licence). |
| Architecture | LoRA, rank 128, alpha 128, dropout 0, on q, k, v, o, gate, up and down projections of all 60 text layers |
| Trainable parameters | 979M |
| Question types | choice, noul (yes/no, returns P(yes)), score (ordered levels) |
| Memory | About 83 GB in bf16, including PyTorch cache. One 96 GB GPU, or two GPUs with the split option. |
Limits
- 32,768 tokens per prompt row (state plus one question).
- 255 options per question. Larger requests are refused as
Unsupported. - The card gives no per-request token limit. The plan sets 64k tokens per request for the API (see limits and errors).
- The card says no public request reached these limits.
Calibration
- Fixed temperature 8.5495, the training temperature.
- Temperature changes probabilities, not the winning option.
Benchmarks (self-reported)
| Measure | Score |
|---|---|
| Decision Index Full score (leaderboard edition 0.3.2, board data of 2026-10-14) | 63.8 |
| Public index (own run, 0.3 suite) | 62.52 (raw 71.08, breadth 61.55) |
Coverage: 140,178 requests over 43 benchmarks. The Full score is the leaderboard figure. It weights public benchmarks at 20%, private tests of the same skills at 50% and private tasks from new domains at 30%.
Latency (measured by the authors)
On 1 x RTX PRO 6000 Blackwell Server Edition, single process, eager, one request at a time, over 400 random public Decision Index 0.3 requests:
- Median 163 ms, mean 644 ms, p80 675 ms.
- The leaderboard lists a median of 253 ms on the same class of GPU.
Latency here is pure GPU compute. Do not compare it with end-to-end HTTPS latency from other providers.
Training
- 190,314 decision questions (76.1M tokens) over 155 tracks.
- 2,742 steps on 2 x NVIDIA B200, about 12 hours.
- Training files are not published.
Requirements
decider-ai1.6.0 (install with--no-deps). The licence of this package is not confirmed. The Sieve inference engine plans to replace it with its own prompt builder.sieve-decisionsfrom GitHub, anddecision-indexat commit62d2f51.- torch 2.8.0, transformers 5.17.0, peft 0.21.0, accelerate 1.15.0, huggingface_hub 1.32.0.
Intended use and limits
The card has no intended-use section. It lists the option, token and memory limits above. It does not state data retention.
Licence
- Adapter and code: Apache-2.0.
- Base model: Gemma terms. Customers who use this model, including through Vault, accept those terms. The Vault container ships the base licence for each model ID (see Vault install).
Data retention
The model keeps nothing between calls. Through the hosted API, Sieve does not store prompts, state or answers (see data retention). Vault customers run the model in their own cloud, so Sieve never sees their content.
Install and call
Through System One: "model": "sieve-31b". See the quickstart.
Next steps
Last updated
Sieve-27B-v2
Model card for Sieve-27B-v2, a large fully fine-tuned decision model with per-category calibration. It returns probabilities and does not generate text.
Sieve-Vision-4B
Model card for Sieve-Vision-4B, a vision decision model for multiple-choice questions about images. It returns probabilities and does not generate text.