echoxvf/SingVL_S1_4B-DI-results

SingVL S1 4B public evaluation

This repository contains the complete Public Decision Index 0.3 results for inclusionAI/SingVL_S1_4B and the inference adapter used to reproduce them. We request evaluation on the private Text and Vision suites.

The model revision is 93f15cd61c1cf417d9cb2631f6be2df13510829e. Its two weight-shard SHA-256 values and core inference source match the evaluated checkpoint and implementation.

Panel	Public index	Requests	Successful
Text	69.76	140178	140178
Vision	73.74949871316771	12210	12210

The run directories contain the original per-request results in lossless gzip form, scores and completion records. Text includes its per-benchmark summary and index. audit/ contains request and image integrity checks, weight provenance and interface validation.

Inference

SingVL S1 4B uses Transformers with prefix-bidirectional full-attention layers and causal GDN layers. Each question is scored by one complete forward pass at the last valid input position. KV caching is disabled. Every supplied candidate receives a probability.

The full public run used the native prediction loop in code/inference/benchmark.py. The inference.engine:D3Engine adapter wraps the same SingVLS1.predict implementation. The recorded interface verification compares it with saved full-run outputs on all 43 Text and 11 Vision benchmarks. The HTTP service supports /v1/systemone for both modalities.

The source run used PPU2 with 16 independent PPU workers, BF16, PyTorch 2.10.0 and Transformers 5.13.1. The kit commit is f5a6e4fd85fed05e77878dfc4966d0568bfb45ec. Private-suite evaluation and single-RTX-PRO-6000 latency qualification are requested from the maintainers.

Reproduction

Download the model at the pinned revision into model/ and this results repository into results/. Install the model requirements and the Decision Index kit at the recorded commit. Build and verify the Public 0.3 suites using that kit.

The following commands run from results/code/. Replace the absolute paths with the downloaded model and verified suites.

python -m decision_index pipeline --edition 0.3 \
  --suite-dir /path/to/text-public-0.3 \
  --engine inference.engine:D3Engine \
  --model /path/to/model --option device=cuda:0 \
  --out /path/to/new-text-run

python -m decision_index vision pipeline \
  --dir /path/to/vision \
  --engine inference.engine:D3Engine \
  --model /path/to/model --option device=cuda:0 \
  --out /path/to/new-vision-run

The engine declares up to four images, up to 256 choice candidates, 65536 input tokens for text and 10000 for inputs with images. Oversized inputs raise Unsupported. Image processing uses native aspect-preserving, patch-grid-aligned RGB bicubic resizing with a 1638400-pixel per-image cap. The audited run reached a maximum of 1638400 pixels per image and 3200 image tokens per request.

HTTP usage and the choice, noul and expected-value score formats are documented in the model repository and code/inference/README.md.

Integrity

File	SHA-256 of uncompressed results
runs/singvl-s1-4b-text/results.jsonl.gz	875e8e87f64005cd48effba7e905354ccd8060eb967102b469b734c70cdadaef
runs/singvl-s1-4b-vision/results.jsonl.gz	41d0e9eac481f74a59ffeb36aa15b1c8a4121d69afeee38fed37359aebdd395d

The original public benchmark images and questions are obtained through the kit's pinned upstream sources. This repository supplies model predictions and evaluation metadata.