@kaggle

Kaggle is the largest global AI community of developers, researchers, and enthusiasts who compete, collaborate, and benchmark what's next in AI.

 San Francisco
kaggle.com
Joined October 2009
  • Tweets5.7K
  • Following292
  • Followers32K
1.4KMedia
Export your AutoScientist open weights to Kaggle 💻 Can't wait to see what the community builds and shares!
adaption
@adaption_ai
Sep 22
Proud to support 33M+ Kaggle builders. 🎉 AutoScientist open weights now export directly to @kaggle.
12
96
12,252
Finish the sentence: "My Kaggle workflow isn't complete without..."
13
1
46
11,860
Post-train an open model into a coding agent that accelerates developer workflows on everyday hardware. The Gemma 4 Developer Agent Competition, hosted by @googlegemma and Kaggle, is live! 💰 Total Prize Pool: $100,000 ⏰ Entry Deadline: November 25, 2026
14
95
13
1,249
72,032
Join and learn more here: kaggle.com/competitions/gemm…

Google - The Gemma 4 Developer Agent Competition

Accelerate research in autonomous coding agents

kaggle.com
3
1
39
6,350
How well can frontier models actually code human-preferred motion graphics? 🎬⚡️ We’re excited to introduce @HeyGen's Code2Video Bench on Kaggle, a standardized benchmark measuring how well models can generate deterministic, high-quality @HyperFrames_ code from natural language creative briefs. 💡 The Problem: Static image benchmarks and traditional video generation evaluations don’t measure whether code generation models can follow detailed creative briefs to programmatically control timing, animation sequences, and asset placement over time. 🛠️ How it Works: 1. Creative Brief: Models receive commercial-grade motion design prompts (example: “Create a 6.8s Figma-style product hook that zooms out from a blooming vector flower into a 3-card web page and live timeline editor”). 2. Code to MP4: The model generates HTML/CSS/JS, deterministically rendered into output.mp4 via the open-source Hyperframes engine in isolated Kaggle Harbor sandboxes. 3. The Auto-Judge: HeyGen’s in-house model architecture, trained on human preferences across a 5-axis evaluation framework, evaluates the rendered MP4. Give it the original prompt and two candidate videos, and it returns the probability a human would prefer Video A over Video B. It does this once per axis — five independent verdicts, not one blended score. Here’s what we’ve learned so far 🔽
A leaderboard graphic titled "HeyGen benchmark: Code2Video," displaying ELO scores from an AI judge grading each model's generated video. The graphic ranks 16 AI models by their ELO scores, represented by horizontal bar charts. Rank 1, GPT-5.5, is highlighted in a blue outlined container with a dark active bar. The full list of ranked models and ELO scores: 1. GPT-5.5: 1574.5 (Highlighted) 2. GPT-6 Astra: 1566.3 3. GPT-5.6 Sol: 1548.7 4. Claude Fable 5.1: 1548.5 5. Claude Opus 5: 1543.0 6. Qwen 3.8 Max: 1539.3 7. GLM-5.3: 1504.2 8. Kimi K3: 1493.7 9. Gemini 3.7 Flash: 1484.3 10. Deepseek V4 Flash: 1484.1 11. GLM-5.3-Flash: 1472.2 12. Gemini 3.1 Pro Preview: 1468.7 13. Grok 4.6: 1464.1 14. GLM-5.2: 1453.2 15. Gemini 3.8 Flash: 1446.1 16. Deepseek V4 Pro: 1409.2 Source at the bottom: HeyGen Code2Video (kaggle.com/benchmarks/heygen/code2video)

ALT A leaderboard graphic titled "HeyGen benchmark: Code2Video," displaying ELO scores from an AI judge grading each model's generated video. The graphic ranks 16 AI models by their ELO scores, represented by horizontal bar charts. Rank 1, GPT-5.5, is highlighted in a blue outlined container with a dark active bar. The full list of ranked models and ELO scores: 1. GPT-5.5: 1574.5 (Highlighted) 2. GPT-6 Astra: 1566.3 3. GPT-5.6 Sol: 1548.7 4. Claude Fable 5.1: 1548.5 5. Claude Opus 5: 1543.0 6. Qwen 3.8 Max: 1539.3 7. GLM-5.3: 1504.2 8. Kimi K3: 1493.7 9. Gemini 3.7 Flash: 1484.3 10. Deepseek V4 Flash: 1484.1 11. GLM-5.3-Flash: 1472.2 12. Gemini 3.1 Pro Preview: 1468.7 13. Grok 4.6: 1464.1 14. GLM-5.2: 1453.2 15. Gemini 3.8 Flash: 1446.1 16. Deepseek V4 Pro: 1409.2 Source at the bottom: HeyGen Code2Video (kaggle.com/benchmarks/heygen/code2video)

12
11
1
102
11,689
more replies
Motion is the weakest axis across every model: • Broken timing: Models hit every beat in the brief, but elements fire before previous animations settle. • Mid-flight collisions: Text is laid out for the final frame, crashing along the path it travels. • Off-by-one reveals: Even #2 GPT-6-Astra slips on typewriter reveals, stopping one character short ("Make it your own" holds on screen as "Make it your owr").
2
3
1,615
When unsure how to fill a frame, models play defensive CSS: small type, thin lines, faint backgrounds, wide margins. Ask for a world map filling 2/3 of the canvas, and models render a tiny, quiet version in the center. Everything compiles and stays legible, but nothing has presence.
3
1,506
Can clever engineering on a small budget still beat massive compute on a Kaggle leaderboard, or has scale officially won?
17
6
3
181
23,591
Mass spectrometry detects thousands of molecules in nature, but most remain unidentified. 🌿 Join the Enveda CASMI 2026 - Molecule ID From Mass Spectra Competition, hosted by @enveda and Kaggle, to build machine learning models that predict 2-D chemical structures from LC-MS/MS spectra. • Prize Pool: $50,000 • Entry Deadline: December 7, 2026
5
7
60
12,734
Your solution could help researchers discover new medicines and identify disease biomarkers. Learn more here: kaggle.com/competitions/enve…

Enveda CASMI 2026 - Molecule ID From Mass Spectra

Predict 2D chemical structures of small molecules detected in complex biological extracts

kaggle.com
2
5
2,789
We’re sitting down with Orbit Wars competitors across the hardware spectrum for our first community podcast episode. What questions do you have for them? 👇
4
4
19
11,159
Introducing ExtractBench on Kaggle Benchmarks with @llama_index. When AI agents rely on schema-guided extraction before human review, one truncated schedule or invented value becomes a wrong payment or decision. ExtractBench evaluates models in workflows based on real-world documents across industries such as supply chain, healthcare, and finance, by measuring whether the system returns: ➣ Missing fields as null instead of inventing a value ➣ Source evidence for each value ➣ Every record of each repeated structure GPT-5.6 Sol currently leads at 91%.
The image shows a bar chart of the ExtractBench benchmark leaderboard by LlamaIndex at Kaggle. The rankings are: 1. GPT-5.6 Sol - 91.0% 2. GPT-5.6 Terra - 90.0% 3. GPT-5.5 - 89.1% 4. Gemini 3 Flash Preview - 89.0% 5. GPT-5.6 Luna - 89.0% 6. Claude Opus 5 - 88.8% 7. Gemini 3.8 Flash - 87.1% 8. GPT-5.4 Mini - 86.4% 9. Gemini 3.5 Flash - 85.6% 10. Gemini 3.7 Flash - 85.5% 11. Claude Haiku 4.5 - 81.6% 12. Gemma 4 31B IT - 79.9% 13. Gemini 3.5 Flash Lite - 79.4% 14. Gemma 4 26B A4B IT - 77.8% 15. GPT-5.4 Nano - 68.9% 16. Gemini 3.6 Flash - 66.3% 17. DeepSeek V3.1 - 0.0% Source: ExtractBench Leaderboard (kaggle.com/benchmarks/llamaindex-org/extractbench-leaderboard)

ALT The image shows a bar chart of the ExtractBench benchmark leaderboard by LlamaIndex at Kaggle. The rankings are: 1. GPT-5.6 Sol - 91.0% 2. GPT-5.6 Terra - 90.0% 3. GPT-5.5 - 89.1% 4. Gemini 3 Flash Preview - 89.0% 5. GPT-5.6 Luna - 89.0% 6. Claude Opus 5 - 88.8% 7. Gemini 3.8 Flash - 87.1% 8. GPT-5.4 Mini - 86.4% 9. Gemini 3.5 Flash - 85.6% 10. Gemini 3.7 Flash - 85.5% 11. Claude Haiku 4.5 - 81.6% 12. Gemma 4 31B IT - 79.9% 13. Gemini 3.5 Flash Lite - 79.4% 14. Gemma 4 26B A4B IT - 77.8% 15. GPT-5.4 Nano - 68.9% 16. Gemini 3.6 Flash - 66.3% 17. DeepSeek V3.1 - 0.0% Source: ExtractBench Leaderboard (kaggle.com/benchmarks/llamaindex-org/extractbench-leaderboard)

6
12
2
117
14,132
Explore the full leaderboard & benchmark details: kaggle.com/benchmarks/llamai…

ExtractBench Leaderboard | Kaggle

A Benchmark for Schema-Guided Enterprise Document Extraction

kaggle.com
3
2,671
Ready to build models that support knee MRI interpretation? 🩻🦵 Check out this starter notebook by Pilkwang Kim. It is a great starting point to read the DICOM acquisitions, sample and normalise the MRI slices, and turn a pretrained vision backbone into twelve abnormality predictions. kaggle.com/code/pilkwang/rsn…

RSNA Knee baseline v1

Explore and run AI code with Kaggle Notebooks | Using data from multiple data sources

kaggle.com
2
21
3
87
13,456
What's one thing you learned on Kaggle? 🌟
26
1
5
73
17,601
Today, we’re excited to launch Adversarial Customer Service on Kaggle Benchmarks, in partnership with @GertLabs. This benchmark is a two-sided security game: one model plays a bank's support agent holding customer records and a verification policy, the other plays a caller who is secretly either the real customer or an identity thief. The agent has to work out which, from the conversation alone, and then either help or refuse. Explore the leaderboard here: kaggle.com/benchmarks/gert-l…
The image shows the leaderboard for the Adversarial Customer Service benchmark by Gert Labs and Kaggle. It signals Claude Opus 4.8 and Gemini 3.6 Flash in the first place, followed by Gemini 3.5 Flash in the second place, and Gemini 3.5 Flash in the third one. You can find the source at the bottom of the picture, the URL in the post.

ALT The image shows the leaderboard for the Adversarial Customer Service benchmark by Gert Labs and Kaggle. It signals Claude Opus 4.8 and Gemini 3.6 Flash in the first place, followed by Gemini 3.5 Flash in the second place, and Gemini 3.5 Flash in the third one. You can find the source at the bottom of the picture, the URL in the post.

7
15
4
131
19,227
I knew I belonged on Kaggle when... (finish the sentence) 👇
14
1
42
11,981
Load more