KnowLine-9B-Gen2 / INFERENCE.md
PEScn's picture
Release KnowLine-9B-Gen2
e285bd4 verified
|
Raw History Blame Contribute Delete
3.93 kB

KnowLine-9B-Gen2: inference and Decision Index reproduction

Model and readout

  • Base model: Qwen/Qwen3.5-9B.
  • bf16 weights, approximately 19.3 GB.
  • Readout temperature T = 1.63 is folded into the final language-model norm.
  • The front-end temperature is 1. temperatures.json supplies {"score": 0.6135}, so score questions use an effective temperature of approximately 1.0; choice and yes/no questions retain T = 1.63.
  • Each question is scored in one prefill by reading option-label probabilities, with no generated reasoning.
  • Text-only requests whose one-pass length is at most 3,000 tokens are read twice. Longer requests and image requests are read once. Question headers and referenced-fields layout are disabled.
  • Limits: 64 questions per request and 255 options per question. Requests are not truncated.

Software

The evaluation used SGLang 0.5.21, Transformers 5.12.1, Torch 2.13.0 and requests 2.34.2.

Keep serve_knowline.sh, knowline_server.py, knowline_layout.py, knowline_engine.py and temperatures.json together in the model directory.

Start the server

From the downloaded model directory:

pip install "sglang==0.5.21" "transformers==5.12.1" "requests==2.34.2"
sha256sum -c SHA256SUMS
MEM=0.90 bash serve_knowline.sh . 0 8080

The launcher starts a bf16 SGLang server on port 9080 and the /v1/systemone front end on port 8080. It uses one GPU, tensor parallel size 1, an fp32 LM head, 16 front-end workers and the model's temperatures.json. The box prompt and coordinate settings are locate / xyxy. The SGLang launcher also uses --mamba-radix-cache-strategy extra_buffer.

MEM=0.90 matches the full public evaluation. Without an override, the launcher defaults to 0.85. HOST defaults to 127.0.0.1.

curl -s http://127.0.0.1:8080/v1/systemone -H 'Content-Type: application/json' -d '{
  "model": "m",
  "state": "Customer: my order arrived broken, I want my money back.",
  "questions": {"refund": {"type": "noul", "instructions": "Should the agent offer a refund?"}}
}'

Reproduce the public evaluation

The official Decision Index kit is pinned to 62d2f51de34a2de64906345b6bc3e98e27ff55c7. Rebuild the public 0.3 suite locally following the kit instructions; the benchmark inputs are not distributed with this model.

With the model server running and the rebuilt suite-0.3 directory available in the kit checkout:

git clone https://github.com/apolinario/decision-index
cd decision-index
git checkout 62d2f51de34a2de64906345b6bc3e98e27ff55c7
pip install -e .
python -m decision_index pipeline --edition 0.3 --engine http \
  --option base_url=http://127.0.0.1:8080 --option model=m \
  --out runs/KnowLine-9B-Gen2

The submitted run used two NVIDIA H20 96 GB GPUs, one identical server per GPU, with 16 round-robin client shards in total. It was a fresh full run with no results carried over from another model or run. Every request succeeded in the first pass; no retries were needed.

  • Public index: 69.00; raw index: 76.40.
  • Requests: 140,620; scoreable: 140,178 after the 442 official exclusions.
  • All scoreable requests were answered; no errors or unsupported requests; complete: true.
  • Runner SHA256: f4c9f36c4c557f537f1d8184db7bf00c5a84d9f4f0aa67f8c06198cead25818d.
  • Results: PelaAI/KnowLine-9B-Gen2-decision-index.

To re-score the downloaded results with the same suite:

python -m decision_index score --edition 0.3 \
  --suite-dir suite-0.3 --results /path/to/results.jsonl.gz \
  --engine http --out rescore

SHA256SUMS records the model package's file hashes. The public index is separate from the private Full score, which is run by the Decision Index maintainers.