加载中…
加载中…
Structured decisions. Measured performance.
AutoTrust-JEV-27B v1.1 is a text decision model that assigns probabilities to the answer options supplied with each question. This card reports its measured performance on the Decision Index 0.3 public suite.
Public index: 67.31 | 37 indexed benchmarks | 140,178 completed scoreable requests
This is a complete local evaluation using the official evaluation kit, self-reported pending maintainer verification. No private or Full score is reported.
The clearest strengths in this evaluation are tool selection, intent classification, and everyday automation decisions. Tools & Automation is the highest-scoring area at 85.19 skill points.
| Capability | Benchmark | This Model |
|---|---|---|
| Intent classification, including out-of-scope detection | CLINC150+OOS | 97.06 |
| Tool selection | BFCL | 96.25 |
| Banking intent classification | BANKING77 | 93.63 |
| Appliance-control decisions | Home appliance simulator | 89.77 |
| API-use decisions | API-Bank | 87.36 |
| Claim verification | HoVer | 78.24 |
| Clinical textual inference | NLI4CT | 75.56 |
| Contract inference | ContractNLI | 74.80 |
These are scores on the Decision Index task adaptations, not claims of perfect real-world reliability or general deployment readiness.
The skill score accounts for each task's chance level: zero represents chance-level performance and 100 represents perfect performance. The following numbers are copied from the official scorer's output. The public index is not a simple average of the rounded values in this table.
| Area | Indexed benchmarks | This Model |
|---|---|---|
| Tools & Automation | 5 | 85.19 |
| Language Understanding | 9 | 73.38 |
| Retrieval & Classification | 6 | 66.97 |
| Knowledge & Reasoning | 11 | 55.16 |
| Arts & Human Taste | 6 | 51.08 |
| Public index | 37 | 67.31 |
The scorer also reports a raw index of 75.62. Raw index and skill index are different metrics and should not be interchanged.
The tables compare AutoTrust-JEV-27B v1.1, JEV-27B, and Jev 1.14 (jev-1.14.0). Each cell is the benchmark's task score multiplied by 100: accuracy, F1 or nDCG according to the task, not the chance-adjusted skill index. Higher is better. Bold marks the best available score in a directly comparable row.
Our column comes from the complete local evaluation; the two reference columns are the published lab results, not new runs performed by us. All three use the same pinned base public corpus. Additional task-specific protocols must also match: GSM8K is not directly compared, because our saved result and the published two-track result use different configurations. A dash means no verified like-for-like value, not zero.
| Benchmark | This Model | JEV-27B | Jev 1.14 |
|---|---|---|---|
| WinoGrande | 95.26 | 82.56 | 91.08 |
| GSM8K (different tracks; not compared) | 84.99 | - | - |
| BBH fixed-option tasks | 86.05 | 78.94 | 93.19 |
| CRUXEval | 85.44 | 73.68 | 74.74 |
| MMLU-Pro | 73.65 | 61.53 | 82.13 |
| CLadder | 81.02 | 70.50 | 88.58 |
| MuSR | 70.61 | 62.37 | 72.61 |
| GPQA Diamond | 54.59 | 44.90 | 75.51 |
| SATA-Bench | 37.88 | 33.64 | 27.88 |
| ChessBench | 21.20 | 16.50 | 16.88 |
| HLE | 16.57 | 6.99 | 19.36 |
| Benchmark | This Model | JEV-27B | Jev 1.14 |
|---|---|---|---|
| HellaSwag | 98.22 | 92.49 | 94.34 |
| FinEntity | 94.47 | 83.35 | 87.68 |
| NLI4CT | 87.44 | 84.27 | 85.12 |
| ContractNLI | 82.58 | 75.04 | 70.76 |
| ANLI | 79.90 | 77.20 | 74.27 |
| VAST | 79.38 | 54.76 | 73.65 |
| RAGTruth response-level hallucination | 83.79 | 80.88 | 75.12 |
| iSarcasmEval | 65.87 | 41.98 | 53.27 |
| ACOS | 56.39 | 19.48 | 33.14 |
| Benchmark | This Model | JEV-27B | Jev 1.14 |
|---|---|---|---|
| CLINC150+OOS | 97.08 | 84.62 | 90.28 |
| BANKING77 | 93.71 | 74.20 | 79.94 |
| HoVer claim verification | 89.12 | 73.90 | 76.75 |
| PhishNChips phishing decisions | 72.00 | 61.90 | 83.75 |
| BRIGHT | 49.19 | 48.44 | 47.05 |
| Amazon ESCI | 52.11 | 54.56 | 57.18 |
| Benchmark | This Model | JEV-27B | Jev 1.14 |
|---|---|---|---|
| BFCL | 97.23 | 95.99 | 96.40 |
| Home appliance simulator | 89.77 | 52.27 | 45.45 |
| API-Bank | 87.60 | 82.87 | 86.61 |
| When2Call MCQ | 89.54 | 78.53 | 81.65 |
| ToolRet | 68.71 | 65.73 | 65.05 |
| Benchmark | This Model | JEV-27B | Jev 1.14 |
|---|---|---|---|
| BPoMP | 96.52 | 92.94 | 90.41 |
| New Yorker caption matching | 82.95 | 68.56 | 72.54 |
| POP909-CL | 61.98 | 39.45 | 15.96 |
| cfcolor | 64.84 | 64.14 | 65.04 |
| Humicroedit | 63.20 | 61.15 | 61.76 |
| Habermas Machine | 42.78 | 40.39 | 44.57 |
jev-1.14.0), lab run dated 2026-10-14. Shared corpus SHA256: b2b56d6fb636837ca469e689087bdbf373dda8de7638aa2da6793e6eda0792d5.| Item | Setting |
|---|---|
| Evaluation | Decision Index 0.3 public suite |
| Completion | complete: true; 140,178 completed scoreable requests |
| Benchmarks included in the index | 37, across five areas |
| Official evaluation kit | Commit f5a6e4f |
| Serving runtime | vLLM 0.31.0 |
| Local hardware | One NVIDIA B300 |
| Concurrency | One server, 16 concurrent clients |
| Reported output | Option probabilities; no generated chain of thought |
| Result status | Self-reported local result; pending maintainer verification |
The saved scorer output reports a median of 64.8 ms, a mean of 178.8 ms, and a p95 of 800.5 ms per request. Requests can contain different numbers of questions. These are local concurrent-run timings, not standardized leaderboard hardware measurements or a latency guarantee for deployment.
Evaluation results, scores and benchmark summaries are archived in the evaluation results dataset, under runs/autotrust-jev-27b-v1.1.
The model is served with vLLM and queried through the decision engine shipped in this repository.
pip install vllm==0.31.0 transformers==5.17.0 "startlux-decide @ git+https://github.com/StartLuxLabs/StartLux-Decision@cb1d11f1ac9e9b43b95ef3a6468fa29b607c40fd"
hf download autotrust/AutoTrust-JEV-27B-v1.1 --local-dir AutoTrust-JEV-27B-v1.1
bash AutoTrust-JEV-27B-v1.1/serve_vllm.sh AutoTrust-JEV-27B-v1.1
serve_vllm.sh turns on the StartLux prefix-bidirectional plugin (STARTLUX_PREFIX_BIDIR=1), so the model reads the whole prompt bidirectionally as it was trained, and sets the flags the results depend on: no prefix caching, no chunked prefill, FlashAttention, processed log-probabilities (up to 255). On an 80 GB card, serve with MAX_LEN=32768.
The vLLM endpoint by itself is not a decision API. jev_vllm_engine.py renders each question in the model's prompt format (jev_format.py), requests one token restricted to the option codes, and turns the returned log-probabilities into option probabilities. With the Decision Index kit installed, run it as an engine:
PYTHONPATH=AutoTrust-JEV-27B-v1.1 python -m decision_index run --engine jev_vllm_engine:JevVllmEngine \
--option model=AutoTrust-JEV-27B-v1.1 --option url=http://127.0.0.1:8090/v1 --out runs/autotrust-jev-27b-v1.1
Supported question types are choice and yes/no; inputs are text only.
Apache-2.0. AutoTrust-JEV-27B-v1.1 is released under the Apache License 2.0 (see LICENSE). This repository contains the fine-tuned AutoTrust-JEV-27B-v1.1 model weights and its decision readout head. Existing upstream attribution and license notices are retained in NOTICE.base and LICENSE.base-apache-2.0.
This license statement applies to this model release; it does not change the licenses or usage restrictions of third-party datasets.
AutoTrust Labs has an official homepage on ModelScope (魔搭社区, Alibaba's model community):
https://www.modelscope.ai/organization/AutoTrust-Labs
A ModelScope mirror for AutoTrust-JEV-27B-v1.1 is not currently confirmed. The model-specific link will be added once that release is available there.