Datasets:

JobOpportunitiesAPI
/

joa-extraction-arena

+# JOA Arena Terms
+Version 1, October 2026. Licensor: Loukas Tzekos, trading as Job Opportunities API (JOA), sole proprietorship, Didaskalisis Papathanasiou Vas. 79, 54629 Thessaloniki, Greece, VAT EL117613696. General contact: hello@jobopportunitiesapi.org. Legal matters: luca@tzekos.eu. Phone: +30 2311 113 603.
+## 1. Scores and documentation
+The files data/scoreboard.csv and data/per_field.csv and the dataset card (README.md) are licensed under Creative Commons Attribution 4.0 International (CC BY 4.0). Credit: "Job Opportunities API (JOA), JOA Job-Posting Extraction Arena".
+## 2. Reference labels
+The file data/gold_labels.jsonl ("the Labels") is made available on these terms:
+1. You may use, copy and share the Labels to evaluate, test and benchmark models and systems, and to publish the results, for any purpose, including commercial purposes.
+2. You may not use the Labels, in whole or in part, to train, fine-tune or distil a machine-learning model, or as a reward or preference signal for one.
+3. If you share the Labels, share them unchanged, with these terms and the canary string, and credit "Job Opportunities API (JOA)".
+4. The listing URLs point to pages owned by third parties. These terms grant no rights over those pages or their content.
+5. The Labels are provided as is, without warranty of accuracy or fitness for any purpose. To the extent permitted by law, the licensor is not liable for any use of them.
+6. We remove a listing on request from the employer or from a person concerned. Please use the current version of the file.

README.md ADDED Viewed

	@@ -0,0 +1,115 @@

+---
+pretty_name: JOA Job-Posting Extraction Arena
+license: other
+license_name: joa-arena-terms
+license_link: LICENSE.md
+language:
+- en
+size_categories:
+- n<1K
+task_categories:
+- text-generation
+tags:
+- information-extraction
+- job-postings
+- benchmark
+- llm-evaluation
+- llm-as-a-judge
+configs:
+- config_name: scoreboard
+  data_files: data/scoreboard.csv
+- config_name: per_field
+  data_files: data/per_field.csv
+- config_name: gold_labels
+  data_files: data/gold_labels.jsonl
+---
+# JOA Job-Posting Extraction Arena
+*Disclosure: Job Opportunities API (JOA, jobopportunitiesapi.org) is an independent data business that sells API access to job-posting data. AI helped run the experiments, check the numbers and draft this text; Loukas (Luca) Tzekos is editorially responsible. Contact: hello@jobopportunitiesapi.org.*
+A benchmark for one narrow task: reading a real job posting and filling 11 structured fields. Job Opportunities API (JOA) built it to choose and train its own extraction models. This repository holds the results for 84 models (68 open-weight models, 13 quantised variants and JOA's 3 fine-tuned models), the per-field scores, the grading method and the reference labels for the listings we are allowed to share (221 of the 250).
+## The task
+Each model receives the text of one job posting, as stored by JOA, plus any structured data found on the page, and must return 11 fields:
+| Field | Meaning |
+|---|---|
+| title | The posting title as published |
+| company | The hiring organisation as named in the posting |
+| location | Where the job is, as stated |
+| salary_currency | ISO currency code of the stated pay |
+| salary_min | Lower bound of the stated pay (equal to salary_max when one figure is given) |
+| salary_max | Upper bound of the stated pay |
+| salary_period | hour, day, week, month or year |
+| seniority | Entry Level (0 years), Entry Level (1 year), Mid (2-3 years), Senior (4+ years), Lead, Manager, Director or Executive |
+| remote_type | onsite, hybrid or remote |
+| posted_date | Date the posting was published (ISO 8601) |
+| is_agency | true when the poster is a staffing agency or third party rather than the employer |
+A field the posting does not state should be left empty. Empty is a valid answer.
+## How it was run
+- **Listings:** 250 real, open job postings sampled from JOA in September 2026, from employers' own recruiting systems and careers pages.
+- **Same input for everyone:** every model received the same prefetched page text and structured data, with a completion budget of up to 24,000 tokens. A reply cut off mid-thought counts as no answer.
+- **Reference labels:** built in September 2026 with Claude Opus agents that read JOA's stored text and the live page, with a confidence level per field. Each cell also records whether its value could be found in the input the models were shown.
+- **Two judges:** each answer was graded field by field by two open models, Kimi-Linear-48B-A3B-Instruct and Seed-OSS-36B-Instruct, against the reference. A cell's score is the mean of the two grades. Both judges also appear as contestants in the table.
+- **Scoring rules:** rule A (the main score) leaves out cells whose reference value appears nowhere in the model's input and that almost no model got right; a missing answer where the value could be found scores 0. Rule B (strict) scores every missing answer as 0. Rule C (lenient) leaves out every "no answer, uncertain reference" cell. In the run of 25 September 2026, rankings under the three rules correlated at 0.99.
+- **Uncertainty:** 95% ranges come from a paired bootstrap over listings (1,000 resamples). Re-running the same model at temperature 0 moved its score by up to 1.5 points, so treat smaller differences as noise.
+- **Dates:** stock and quantised models graded 23–29 September 2026, JOA Advance 29 September 2026, JOA Small and JOA Deep 2 October 2026. None of the 250 listings is in the fine-tunes' training data.
+## Top of the scoreboard (rule A, 84 models, as of 2 October 2026)
+| Rank | Model | Score | 95% range | Answered | Median output tokens |
+|---|---|---|---|---|---|
+| 1 | JOA Small (fine-tuned ibm-granite/granite-4.2-3b) | 93.3 | 92.2–94.4 | 100% | 117 |
+| 2 | JOA Advance (fine-tuned ibm-granite/granite-3.3-8b-instruct) | 92.7 | 91.6–93.9 | 100% | 123 |
+| 3 | JOA Deep (fine-tuned Qwen/Qwen3.8-27B) | 92.2 | 91.0–93.4 | 100% | 111 |
+| 4 | Qwen/Qwen3.5-27B (the fine-tunes' trainer) | 89.2 | 86.9–91.4 | 98% | 2,946 |
+| 5 | Qwen/Qwen3-235B-A22B-Instruct-2507 | 89.0 | 86.7–91.2 | 100% | 307 |
+| 9 | Qwen/Qwen3.8-27B (stock) | 88.7 | 86.5–90.8 | 100% | 2,517 |
+| 20 | ibm-granite/granite-3.3-8b-instruct (stock) | 86.4 | 84.4–88.2 | 99% | 266 |
+| 71 | ibm-granite/granite-4.2-3b (stock) | 64.5 | 59.7–69.0 | 80% | 4,411 |
+The full table is `data/scoreboard.csv`; per-field scores for every model are in `data/per_field.csv`.
+## Files
+| File | Rows | Content |
+|---|---|---|
+| data/scoreboard.csv | 84 | rank, model, kind (stock, gguf-quantised, fine-tuned), rule A score with 95% range, rule B and C scores, share of listings answered, median output tokens, size in GB for stock checkpoints, statistically tied with #1 |
+| data/per_field.csv | 84 | rule A score per field, rounded to whole points |
+| data/gold_labels.jsonl | 221 | reference labels: arena_id, listing URL, host, the 11 fields, per-field confidence, per-field "findable in model input" flag, difficulty band |
+## What is not included, and why
+- **The posting text.** Job postings are written by employers and remain theirs. We publish the listing URL instead. Many of these postings have since closed, so the exact input cannot be rebuilt from the URLs alone; the scores above are the record of the run.
+- **Listings we do not share.** The reference labels cover 221 of the 250 listings: those from employer-direct sources that JOA is allowed to redistribute. Left out: listings from a public employment service feed we use only for discovery, from job boards and aggregators, and from small employers named after a person. The scores were computed on all 250.
+- **Personal data.** No names, email addresses or phone numbers of contacts. Free-text evidence quotes used while building the labels are left out.
+## Licence
+The scores, per-field tables and this card are licensed CC BY 4.0. The reference labels in `data/gold_labels.jsonl` are provided for evaluation under the JOA Arena Terms in LICENSE.md: use them freely to evaluate, benchmark and publish results, including commercially, but do not use them to train or fine-tune models. The labels were produced with the help of Claude (Anthropic); this notice keeps the set an evaluation set. To help keep it out of training corpora, the files carry this canary string: `JOA-ARENA-CANARY ac45c5b8-da08-4fb3-8e43-0c178dc94806`.
+Model names in the tables are those of their publishers; their mention does not imply endorsement.
+## Citation
+Job Opportunities API (2026). JOA Job-Posting Extraction Arena, scoreboard of 2 October 2026. https://huggingface.co/datasets/JobOpportunitiesAPI/joa-extraction-arena
+## Acknowledgement
+This work used the EuroHPC supercomputer Discoverer+ in Bulgaria, made available by the EuroHPC Joint Undertaking through an AI Factories Playground access allocation (project EHPC-AIF-2026PG01-1124).
+## Removal
+Employers or individuals who want a listing removed: hello@jobopportunitiesapi.org. More about Job Opportunities API (JOA): https://jobopportunitiesapi.org
+## Contact
+- General: hello@jobopportunitiesapi.org
+- Legal matters: luca@tzekos.eu
+- Phone: +30 2311 113 603
+- Job Opportunities API (JOA) is a sole proprietorship of Loukas Tzekos, Didaskalisis Papathanasiou Vas. 79, 54629 Thessaloniki, Greece. VAT EL117613696.

data/gold_labels.jsonl ADDED Viewed

The diff for this file is too large to render. See raw diff

data/per_field.csv ADDED Viewed

	@@ -0,0 +1,85 @@

+model,title,company,location,salary_currency,salary_min,salary_max,salary_period,seniority,remote_type,posted_date,is_agency
+JOA Small (fine-tuned ibm-granite/granite-4.2-3b),99,96,89,93,94,93,96,86,92,89,96
+JOA Advance (fine-tuned ibm-granite/granite-3.3-8b-instruct),99,96,90,93,94,94,95,83,91,84,96
+JOA Deep (fine-tuned Qwen/Qwen3.8-27B),99,96,90,92,93,93,96,86,92,72,96
+Qwen/Qwen3.5-27B,95,90,86,90,93,86,92,82,85,88,93
+Qwen/Qwen3-235B-A22B-Instruct-2507,96,89,86,90,91,87,93,83,95,90,80
+gguf-Qwen3.8-27B-UD-Q5_K_M,96,93,89,90,92,88,84,71,84,92,93
+ornith-ai/Ornith-1.5-35B-A3B,95,91,87,88,91,87,85,87,80,92,92
+gguf-Qwen3.5-27B-Q5_K_M,95,91,86,89,91,83,88,78,88,87,93
+Qwen/Qwen3.8-27B,96,93,89,90,92,88,85,66,85,92,92
+gguf-Qwen3.5-27B-Q5_K_M-kvq4,94,90,86,90,91,84,89,80,87,87,92
+gguf-Ornith-1.5-35B-A3B-IQ3_XXS,95,91,88,90,91,85,85,87,75,92,92
+Qwen/Qwen3-235B-A22B-Instruct-2507-FP8,95,90,86,90,91,86,93,82,93,90,78
+Qwen/QwQ-32B,96,89,86,90,91,89,92,75,89,75,90
+openai/gpt-oss-120b,97,92,88,89,94,91,86,64,77,94,88
+Qwen/Qwen3.5-122B-A10B,93,89,84,89,91,87,87,80,78,90,91
+gguf-Qwen3.8-Flash-Next-UD-Q4_K_XL,96,91,88,89,90,87,76,69,75,92,92
+Qwen/Qwen3.6-27B,97,93,88,90,93,83,91,72,62,92,93
+Qwen/Qwen3-32B,96,90,86,92,92,88,90,63,75,92,89
+ibm-granite/granite-3.3-8b-instruct-FP8,98,91,87,89,88,87,83,74,80,82,87
+ibm-granite/granite-3.3-8b-instruct,97,92,88,88,89,87,82,73,86,78,85
+Qwen/Qwen3.8-Flash-Next,96,91,88,90,91,88,76,67,71,92,93
+gguf-Qwen3.8-27B-UD-IQ3_XXS,96,92,88,87,90,87,78,57,79,92,92
+openai/gpt-oss-20b,93,88,86,86,91,88,86,64,85,90,86
+gguf-Ornith-1.5-35B-A3B-Q4_K_M,93,90,84,85,86,83,80,86,75,88,91
+gguf-granite-3.3-8b-Q4_K_M,97,93,87,88,87,86,78,76,80,80,84
+ByteDance-Seed/Seed-OSS-36B-Instruct,97,91,86,90,92,83,82,63,69,92,91
+Qwen/Qwen3-Next-80B-A3B-Instruct-FP8,93,89,86,87,89,86,89,81,73,77,88
+Qwen/Qwen3-Next-80B-A3B-Instruct,94,90,86,89,89,85,91,80,73,74,87
+gguf-granite-3.3-8b-Q4_K_M-kvq8,97,92,88,88,85,86,78,75,78,79,84
+gguf-granite-3.3-8b-f16,96,91,87,88,89,87,77,73,83,77,84
+gguf-granite-3.3-8b-Q8_0,96,90,87,88,89,87,76,72,82,77,85
+openai/gpt-oss-safeguard-20b,91,88,86,86,88,87,86,66,79,90,87
+gguf-granite-3.3-8b-Q4_K_M-kvq4,97,92,88,88,84,84,74,75,69,82,86
+Qwen/QwQ-32B-Preview,91,87,85,84,85,81,86,78,90,56,89
+Qwen/Qwen3.6-35B-A3B,95,91,87,89,92,86,88,58,56,88,91
+ibm-granite/granite-4.2-8b,94,88,87,86,88,81,71,72,70,91,84
+deepseek-ai/DeepSeek-R1-Distill-Qwen-14B,97,91,86,91,91,90,79,55,71,94,74
+deepseek-ai/DeepSeek-R1-Distill-Qwen-32B,96,91,86,91,91,89,87,55,60,91,85
+stepfun-ai/Step-3.5-Flash,95,89,82,83,89,84,74,51,79,83,90
+Qwen/Qwen3-30B-A3B-Instruct-2507,94,87,85,89,90,87,92,74,53,80,87
+Qwen/Qwen3-30B-A3B-Instruct-2507-FP8,94,87,85,88,89,85,92,73,54,77,88
+ibm-granite/granite-4.2-30b,95,88,83,85,87,82,71,57,85,90,74
+ibm-granite/granite-4.2-30b-fp8,95,88,83,87,88,85,75,57,78,91,75
+ibm-granite/granite-4.2-30b-mxfp4,95,89,86,88,88,83,74,52,79,91,71
+allenai/Olmo-3.1-32B-Instruct,93,86,85,88,89,78,89,77,62,88,72
+Qwen/Qwen3-14B,95,89,86,88,89,87,79,60,73,89,65
+Qwen/Qwen3-4B-Instruct-2507-FP8,95,88,83,90,91,88,92,67,54,82,71
+Qwen/Qwen3-4B-Instruct-2507,94,86,84,89,90,86,90,67,54,81,74
+moonshotai/Kimi-Linear-48B-A3B-Instruct,95,87,86,88,91,80,93,65,52,60,88
+Qwen/Qwen3-30B-A3B,96,89,85,89,88,85,86,49,54,94,73
+ibm-granite/granite-4.1-30b,95,90,86,86,88,81,70,38,51,92,85
+ibm-granite/granite-3.3-2b-instruct,91,87,81,80,79,78,55,68,84,78,73
+zai-org/GLM-4.5-Air,89,84,81,80,86,75,78,67,65,71,83
+Qwen/Qwen3-8B,94,90,85,88,88,81,81,44,56,92,70
+Qwen/Qwen3-8B-FP8,92,88,83,89,90,82,86,44,52,92,72
+Qwen/Qwen3-4B,94,85,82,86,86,83,78,55,66,92,57
+Qwen/Qwen2.5-Coder-32B-Instruct,85,78,77,81,81,77,82,78,65,80,77
+zai-org/GLM-4-9B-0414,98,93,87,90,90,79,89,61,51,94,36
+openbmb/MiniCPM5-2B,96,88,83,88,91,86,76,29,55,93,66
+Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8,90,83,81,90,88,85,88,55,46,81,72
+Qwen/Qwen3-Coder-30B-A3B-Instruct,92,84,83,89,87,84,86,53,47,80,64
+deepseek-ai/DeepSeek-R1-0528-Qwen3-8B,93,90,85,84,89,78,79,28,55,91,60
+zai-org/GLM-4.7-Flash,86,82,76,74,77,75,73,55,55,80,79
+Qwen/Qwen2.5-7B-Instruct,94,89,85,90,91,83,89,64,46,86,20
+baidu/ERNIE-4.5-21B-A3B-Thinking,95,79,80,75,82,78,64,31,41,71,75
+openbmb/MiniCPM5-2B-Midtrain,81,79,75,65,76,73,64,60,60,76,61
+ibm-granite/granite-4.1-3b,86,84,80,77,80,72,48,22,57,81,69
+HuggingFaceTB/SmolLM3-3B,88,82,77,67,77,73,72,19,43,78,76
+gguf-Qwen3.5-27B-UD-IQ3_XXS,72,69,65,71,72,70,73,60,67,71,69
+Qwen/Qwen3-1.7B,96,85,81,68,62,59,44,36,54,79,55
+ibm-granite/granite-4.2-3b,76,69,68,60,66,60,38,37,72,77,65
+deepseek-ai/DeepSeek-R1-Distill-Qwen-7B,94,80,83,61,59,57,55,20,35,67,65
+Qwen/Qwen3.5-9B,62,60,57,67,68,67,68,60,54,65,57
+openbmb/MiniCPM3-4B,98,90,83,62,66,63,54,29,33,91,1
+ibm-granite/granite-4.0-h-1b,85,81,75,68,75,69,18,42,21,81,35
+ibm-granite/granite-4.0-1b,95,90,81,30,30,26,1,67,78,46,17
+ibm-granite/granite-3.1-1b-a400m-instruct,83,75,70,49,45,46,23,48,58,79,42
+ibm-granite/granite-4.1-8b-fp8,66,61,60,49,51,46,39,42,33,42,59
+microsoft/Phi-4-mini-instruct,86,78,73,56,56,53,44,4,36,56,1
+ibm-granite/granite-4.1-8b,59,56,54,43,45,40,34,41,33,35,52
+deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B,71,51,56,36,24,25,7,32,41,29,39
+Qwen/Qwen3.5-4B,45,43,42,33,34,33,31,37,37,39,37
+openbmb/MiniCPM5-1B,40,35,32,21,14,14,3,22,15,32,15
+ibm-granite/granite-3.0-1b-a400m-instruct,30,28,23,6,6,5,2,12,12,14,19

data/scoreboard.csv ADDED Viewed

	@@ -0,0 +1,85 @@

+rank,model,kind,score_ruleA,ci_low,ci_high,strict_B,lenient_C,answered_pct,median_output_tokens,size_GB,tied_with_1
+1,JOA Small (fine-tuned ibm-granite/granite-4.2-3b),fine-tuned,93.3,92.2,94.4,80.6,84.0,100,117,,True
+2,JOA Advance (fine-tuned ibm-granite/granite-3.3-8b-instruct),fine-tuned,92.7,91.6,93.9,80.1,83.9,100,123,,True
+3,JOA Deep (fine-tuned Qwen/Qwen3.8-27B),fine-tuned,92.2,91.0,93.4,79.4,83.1,100,111,,False
+4,Qwen/Qwen3.5-27B,stock,89.2,86.9,91.4,76.6,81.4,98,2946,56,False
+5,Qwen/Qwen3-235B-A22B-Instruct-2507,stock,89.0,86.7,91.2,76.9,80.2,100,307,470,False
+6,gguf-Qwen3.8-27B-UD-Q5_K_M,gguf-quantised,88.8,86.7,90.8,76.3,81.9,100,2388,,False
+7,ornith-ai/Ornith-1.5-35B-A3B,stock,88.8,86.3,91.1,76.5,80.9,99,2000,72,False
+8,gguf-Qwen3.5-27B-Q5_K_M,gguf-quantised,88.8,86.4,91.1,76.3,81.2,99,1558,,False
+9,Qwen/Qwen3.8-27B,stock,88.7,86.5,90.8,76.1,81.9,100,2517,56,False
+10,gguf-Qwen3.5-27B-Q5_K_M-kvq4,gguf-quantised,88.5,85.9,91.1,76.2,81.0,97,1842,,False
+11,gguf-Ornith-1.5-35B-A3B-IQ3_XXS,gguf-quantised,88.4,86.1,90.7,76.3,80.4,99,2017,,False
+12,Qwen/Qwen3-235B-A22B-Instruct-2507-FP8,stock,88.4,86.1,90.5,76.6,79.9,100,303,236,False
+13,Qwen/QwQ-32B,stock,87.7,85.4,89.9,75.3,79.9,100,995,66,False
+14,openai/gpt-oss-120b,stock,87.3,85.6,89.0,74.9,81.0,100,821,196,False
+15,Qwen/Qwen3.5-122B-A10B,stock,87.0,84.4,89.6,74.8,80.0,97,3938,250,False
+16,gguf-Qwen3.8-Flash-Next-UD-Q4_K_XL,gguf-quantised,86.7,84.3,88.9,74.5,80.4,100,1459,,False
+17,Qwen/Qwen3.6-27B,stock,86.7,84.7,88.6,74.4,79.9,100,2603,56,False
+18,Qwen/Qwen3-32B,stock,86.4,84.5,88.3,74.4,79.7,100,564,66,False
+19,ibm-granite/granite-3.3-8b-instruct-FP8,stock,86.4,84.6,87.9,75.3,78.2,100,285,8,False
+20,ibm-granite/granite-3.3-8b-instruct,stock,86.4,84.4,88.2,75.1,78.1,99,266,16,False
+21,Qwen/Qwen3.8-Flash-Next,stock,86.3,83.9,88.4,74.0,80.2,100,1391,360,False
+22,gguf-Qwen3.8-27B-UD-IQ3_XXS,gguf-quantised,86.1,83.6,88.3,73.8,80.2,99,1868,,False
+23,openai/gpt-oss-20b,stock,86.0,83.6,88.0,73.8,78.9,100,845,41,False
+24,gguf-Ornith-1.5-35B-A3B-Q4_K_M,gguf-quantised,85.9,82.6,88.8,74.0,78.6,97,1768,,False
+25,gguf-granite-3.3-8b-Q4_K_M,gguf-quantised,85.6,83.6,87.3,74.7,77.6,99,253,,False
+26,ByteDance-Seed/Seed-OSS-36B-Instruct,stock,85.4,83.1,87.6,73.3,79.2,100,1650,72,False
+27,Qwen/Qwen3-Next-80B-A3B-Instruct-FP8,stock,85.3,82.6,87.8,73.7,77.8,100,296,82,False
+28,Qwen/Qwen3-Next-80B-A3B-Instruct,stock,85.2,82.5,87.7,73.7,77.9,100,307,163,False
+29,gguf-granite-3.3-8b-Q4_K_M-kvq8,gguf-quantised,85.2,83.4,86.9,74.5,77.1,99,243,,False
+30,gguf-granite-3.3-8b-f16,gguf-quantised,85.1,83.0,87.1,73.9,77.0,98,270,,False
+31,gguf-granite-3.3-8b-Q8_0,gguf-quantised,85.1,83.0,87.0,73.9,77.0,98,264,,False
+32,openai/gpt-oss-safeguard-20b,stock,85.1,82.7,87.3,73.0,78.0,100,691,14,False
+33,gguf-granite-3.3-8b-Q4_K_M-kvq4,gguf-quantised,84.2,82.6,85.8,73.8,76.7,99,258,,False
+34,Qwen/QwQ-32B-Preview,stock,84.0,81.2,86.7,73.1,76.4,98,1351,66,False
+35,Qwen/Qwen3.6-35B-A3B,stock,83.5,81.0,85.8,71.6,77.9,100,2684,72,False
+36,ibm-granite/granite-4.2-8b,stock,83.4,80.6,85.8,71.7,77.0,98,4686,18,False
+37,deepseek-ai/DeepSeek-R1-Distill-Qwen-14B,stock,83.2,81.3,85.2,71.6,76.4,100,573,30,False
+38,deepseek-ai/DeepSeek-R1-Distill-Qwen-32B,stock,83.2,81.0,85.3,71.4,77.1,100,553,66,False
+39,stepfun-ai/Step-3.5-Flash,stock,82.9,80.4,85.4,71.2,77.6,100,4459,399,False
+40,Qwen/Qwen3-30B-A3B-Instruct-2507,stock,82.8,80.4,85.1,71.3,76.2,100,213,61,False
+41,Qwen/Qwen3-30B-A3B-Instruct-2507-FP8,stock,82.6,80.1,84.9,71.1,76.0,100,227,31,False
+42,ibm-granite/granite-4.2-30b,stock,82.3,79.5,84.8,70.6,76.7,100,6973,59,False
+43,ibm-granite/granite-4.2-30b-fp8,stock,82.0,79.4,84.5,70.4,77.0,100,7373,30,False
+44,ibm-granite/granite-4.2-30b-mxfp4,stock,81.9,79.4,84.2,70.3,76.9,100,6520,17,False
+45,allenai/Olmo-3.1-32B-Instruct,stock,81.7,79.1,83.9,70.9,74.8,99,213,64,False
+46,Qwen/Qwen3-14B,stock,81.5,79.0,83.8,70.0,75.6,100,868,30,False
+47,Qwen/Qwen3-4B-Instruct-2507-FP8,stock,80.8,78.4,82.9,69.5,74.4,100,265,5,False
+48,Qwen/Qwen3-4B-Instruct-2507,stock,80.4,77.8,82.8,69.1,74.3,100,257,8,False
+49,moonshotai/Kimi-Linear-48B-A3B-Instruct,stock,80.3,77.8,82.4,69.2,74.3,100,170,98,False
+50,Qwen/Qwen3-30B-A3B,stock,80.0,77.8,82.1,68.7,75.2,100,1013,61,False
+51,ibm-granite/granite-4.1-30b,stock,79.0,76.7,81.3,67.8,74.5,100,224,58,False
+52,ibm-granite/granite-3.3-2b-instruct,stock,78.9,75.8,81.7,69.6,72.0,94,287,5,False
+53,zai-org/GLM-4.5-Air,stock,78.7,75.1,82.0,67.6,74.0,92,1583,221,False
+54,Qwen/Qwen3-8B,stock,78.7,75.8,81.2,67.5,74.1,97,982,16,False
+55,Qwen/Qwen3-8B-FP8,stock,78.4,75.5,81.0,67.3,73.9,95,1025,9,False
+56,Qwen/Qwen3-4B,stock,77.8,74.9,80.5,67.2,72.4,96,838,8,False
+57,Qwen/Qwen2.5-Coder-32B-Instruct,stock,77.7,73.6,81.6,67.1,71.9,100,197,66,False
+58,zai-org/GLM-4-9B-0414,stock,77.2,75.3,79.0,66.6,72.0,100,230,19,False
+59,openbmb/MiniCPM5-2B,stock,76.8,74.8,78.9,65.9,72.7,99,1762,5,False
+60,Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8,stock,76.8,73.9,79.6,65.8,71.8,100,219,31,False
+61,Qwen/Qwen3-Coder-30B-A3B-Instruct,stock,76.1,73.1,78.8,65.3,71.5,100,228,61,False
+62,deepseek-ai/DeepSeek-R1-0528-Qwen3-8B,stock,75.3,73.3,77.3,64.7,71.9,100,2851,16,False
+63,zai-org/GLM-4.7-Flash,stock,74.2,70.0,78.5,63.7,70.1,87,2531,62,False
+64,Qwen/Qwen2.5-7B-Instruct,stock,73.4,71.2,75.5,63.0,68.9,100,240,15,False
+65,baidu/ERNIE-4.5-21B-A3B-Thinking,stock,70.8,68.3,73.3,60.8,67.5,100,3251,44,False
+66,openbmb/MiniCPM5-2B-Midtrain,stock,70.3,67.1,73.5,60.7,65.3,99,1039,5,False
+67,ibm-granite/granite-4.1-3b,stock,70.2,66.5,73.7,60.2,67.5,92,181,7,False
+68,HuggingFaceTB/SmolLM3-3B,stock,69.4,65.8,72.7,59.7,66.5,92,206,6,False
+69,gguf-Qwen3.5-27B-UD-IQ3_XXS,gguf-quantised,68.8,63.7,74.3,59.3,65.6,75,5012,,False
+70,Qwen/Qwen3-1.7B,stock,67.8,65.2,70.4,58.7,64.2,98,849,4,False
+71,ibm-granite/granite-4.2-3b,stock,64.5,59.7,69.0,55.4,61.3,80,4411,7,False
+72,deepseek-ai/DeepSeek-R1-Distill-Qwen-7B,stock,63.8,61.1,66.6,55.1,61.1,99,743,15,False
+73,Qwen/Qwen3.5-9B,stock,61.1,55.4,66.9,52.5,59.0,64,5653,19,False
+74,openbmb/MiniCPM3-4B,stock,61.1,58.4,63.7,52.5,59.0,99,219,8,False
+75,ibm-granite/granite-4.0-h-1b,stock,59.2,55.7,62.4,51.3,56.5,92,218,3,False
+76,ibm-granite/granite-4.0-1b,stock,57.8,55.5,60.3,51.2,54.4,97,105,3,False
+77,ibm-granite/granite-3.1-1b-a400m-instruct,stock,57.4,53.7,61.0,52.1,55.1,82,331,3,False
+78,ibm-granite/granite-4.1-8b-fp8,stock,51.5,46.2,56.6,44.4,50.0,71,253,10,False
+79,microsoft/Phi-4-mini-instruct,stock,50.4,47.2,53.5,43.2,50.1,95,239,8,False
+80,ibm-granite/granite-4.1-8b,stock,46.5,40.8,51.5,40.1,45.5,64,261,18,False
+81,deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B,stock,41.1,37.8,44.4,36.2,40.7,84,660,4,False
+82,Qwen/Qwen3.5-4B,stock,38.5,32.8,44.5,33.2,38.7,47,24000,9,False
+83,openbmb/MiniCPM5-1B,stock,23.9,20.0,28.3,20.7,24.5,45,24000,2,False
+84,ibm-granite/granite-3.0-1b-a400m-instruct,stock,16.3,13.0,19.8,15.6,18.5,44,180,3,False