Decision model
Chinese Decision Pointer 4B
1 question per request
- Accuracy, all 2,198 questions
- 46.7%44.6% to 48.8%, 95% interval; 1,026 right
- Answers kept at 99% confidence
- 51 of 2,1981 wrong; share right 89.7% to 99.7%, 95% interval
- Cost, $ per 1,000 questionsper 1M questions
- $0.025$24.58Basis and price under Cost
- Latency
- 0.035 smedian request, 1 question per request
How we called it
- Kind
- Decision model
- Access
- Decision Index engine (compactor's di_serve.py around the engine its card ships), self-hosted, BF16
- Paid as
- Our own hardware
- Model ID
Chinese-Decision-Pointer-4B- Parameters
- 4B dense, as the maker states it (counted from the released files: 4.21B).language model with its pointer readout, text only; source: counted from the checkpoint's safetensors, quincyqiang/chinese-decision-pointer-4b@889f23e, huggingface.co/quincyqiang/chinese-decision-pointer-4b; read 15 November 2026
- Hardware
- 1 × RTX PRO 6000 Blackwell (96 GB); BF16
- Note
- quincyqiang's released Chinese Decision Pointer 4B (quincyqiang/chinese-decision-pointer-4b at 889f23e), a merged LoRA fine-tune of Qwen3.5-4B that reads every candidate's probability in one forward pass. Its card ships only a Decision Index engine (ChineseDecisionEngine from yanqiangmiffy/chinese-decision at c16450b), so it ran on our hardware behind compactor's di_serve.py, with the card's debias 'reverse' (the original and reversed option order averaged), without causal-conv1d (PyTorch's reference convolution), text only.
- Confidence from
- Option score
- Run
- 1206, 15 November 2026
Settings
- Reasoning
- off
- Request format
- 1 question per request
- Requests at once
- 1
- Temperature
- not recorded
Run record: harness seed, harness commit and prompt version
- Harness seed
- 0
- Harness commit
7bc1f66c76b0- Prompt version
254597b8be16
Results
| Questions | Accuracy | 95% interval | Right |
|---|---|---|---|
| All questions | 46.7% | 44.6% to 48.8% | 1,026 of 2,198 |
| Reading | 70.9% | 67.3% to 74.2% | 465 of 656 |
| Grammar | 36.4% | 34.0% to 38.8% | 561 of 1,542 |
Comparison group: ηθ. Compare with another setup.
Accuracy by topic
English
The Text tab's 547 kept reading questions translated into English, asked with this setup's Text tab settings. Every setup: English reading; against Turkish: English and Turkish.
- Right answers
- 440 of 547, first English run (1209)
- Accuracy
- 80.4%
- Non-hallucination with “I don't know”
- 3.7%: 4 “I don't know” of 108 not right (run 1211)
- Answered with “I don't know”
- 99.3%: 543 of 547
- Error rate with “I don't know”
- 19.2%: 104 of 543 answers
- Cost, $ per 1,000 questionsper 1M questions, first English run
- 0.02727.21from GPU time
- Latency, median request, first English run
- 0.038 s; 1 request at once
- English run
- 1209
- Machine, English runs
- RTX PRO 6000
- Date
- 15 November 2026
- Request format
- 1 question per request
- Reasoning
- off
- Prompt version
47f5776e71db- Translations, SHA-256
bf0d47cd385a93050c65906977e199ef7e45d5848dbbdef2bc8d124f4d5abf68
Confidence
- Answers kept at 99% or more
- 51, of which 1 wrong (share right: 89.7% to 99.7%, 95% interval); 2,147 skipped, the model less than 99% sure
- Calibration gap
- 11.7 points
- AUROC, all questions
- 0.73 0.71 to 0.75
- AUROC, reading
- 0.81 0.78 to 0.85
- AUROC, grammar
- 0.63 0.60 to 0.66
- Brier score
- 0.220
- Accuracy of its surest answers
- most confident fifth: 83.4% right; most confident half: 62.2% right; most confident four fifths: 52.1% right
- Typical score of a right answer
- 68.1%
- Typical score of a wrong answer
- 46.2%
Repeats
We did not repeat this setup.
Latency
- Latency, median request
- 0.035 s (this run)
- Latency, 90th percentile
- 0.043 s
- Requests at once
- 1
- Throughput
- 1688.4 questions per minute of request time
- Request time
- 1.3 minutes, over the requests whose answers were kept
Latency: the time from sending one request to receiving its answer, measured from our computer, network and any retries within the request included. Request time: the time at least one request was in flight, over the requests whose answers were kept. It is not the run's elapsed time: it leaves out requests that failed and any pause before a run was resumed.
Cost
- Cost per 1,000 questionsper 1M questions
- $0.025$24.58
- Whole run
- $0.054
- Basis
- Estimated from GPU time: the run's request time x the GPUs its server held x Runpod's on-demand price for an RTX PRO 6000, $2.49 per GPU-hour (read 5 November 2026). A cloud-equivalent estimate, not measured spending.
- Price
- $2.49 per GPU-hour, 1 GPU, for the run's request time (source, 2026-11-05)
- Input tokens per question
- 728
- Output tokens per question
- 0