Decision model
Verdin-4B-v1.1
1 question per request
- Accuracy, all 2,198 questions
- 46.8%44.7% to 48.9%, 95% interval; 1,028 right
- Answers kept at 99% confidence
- 27 of 2,1980 wrong; share right 87.5% to 100.0%, 95% interval
- Cost, $ per 1,000 questionsper 1M questions
- $0.034$34.21Basis and price under Cost
- Latency
- 0.049 smedian request, 1 question per request
How we called it
- Kind
- Decision model
- Access
- Verdin server (the card's own python -m verdin.server, in the package versions its card lists), self-hosted, BF16
- Paid as
- Our own hardware
- Model ID
Verdin-4B-v1.1- Parameters
- 4B dense, as the maker states it (counted from the released files: 4.21B).language model with the merged LoRA and its pointer head; its card says 4.21B served; source: counted from the checkpoint's safetensors, victorbennett/Verdin-4B-v1.1@45830d2, huggingface.co/victorbennett/Verdin-4B-v1.1; read 15 November 2026
- Hardware
- 1 × RTX PRO 6000 Blackwell (96 GB); BF16
- Note
- victorbennett's released Verdin-4B-v1.1 (victorbennett/Verdin-4B-v1.1 at 45830d2): Qwen3.5-4B with a LoRA fine-tune merged into its weights (the mean of three seeds' LoRA runs) and a pointer head that reads the option-boundary tokens, 4.21B parameters served. It ran on our hardware through its card's own server, without causal-conv1d (its torch convolution path), text only. Weights CC-BY-NC-4.0 (research use).
- Confidence from
- Option score
- Run
- 1244, 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
28d1c0e8e51a- Prompt version
254597b8be16
Results
| Questions | Accuracy | 95% interval | Right |
|---|---|---|---|
| All questions | 46.8% | 44.7% to 48.9% | 1,028 of 2,198 |
| Reading | 71.2% | 67.6% to 74.5% | 467 of 656 |
| Grammar | 36.4% | 34.0% to 38.8% | 561 of 1,542 |
Comparison group: θ. 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
- 450 of 547, first English run (1247)
- Accuracy
- 82.3%
- Non-hallucination with “I don't know”
- 2.1%: 2 “I don't know” of 96 not right (run 1249)
- Answered with “I don't know”
- 99.6%: 545 of 547
- Error rate with “I don't know”
- 17.2%: 94 of 545 answers
- Cost, $ per 1,000 questionsper 1M questions, first English run
- 0.03434.28from GPU time
- Latency, median request, first English run
- 0.052 s; 1 request at once
- English run
- 1247
- 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
- 27, of which 0 wrong (share right: 87.5% to 100.0%, 95% interval); 2,171 skipped, the model less than 99% sure
- Calibration gap
- 9.8 points
- AUROC, all questions
- 0.75 0.73 to 0.77
- AUROC, reading
- 0.83 0.80 to 0.86
- AUROC, grammar
- 0.66 0.63 to 0.69
- Brier score
- 0.207
- Accuracy of its surest answers
- most confident fifth: 86.0% right; most confident half: 65.6% right; most confident four fifths: 52.5% right
- Typical score of a right answer
- 71.4%
- Typical score of a wrong answer
- 40.8%
Repeats
We did not repeat this setup.
Latency
- Latency, median request
- 0.049 s (this run)
- Latency, 90th percentile
- 0.056 s
- Requests at once
- 1
- Throughput
- 1213.0 questions per minute of request time
- Request time
- 1.8 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
- $0.034$34.21
- Whole run
- $0.075
- 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
- 436
- Output tokens per question
- 0