Анализ данных (Data analysis)
Data science, наука о данных. @haarrp - админ РКН: clck.ru/3FmyAp
Show more📈 Analytical overview of Telegram channel Анализ данных (Data analysis)
Channel Анализ данных (Data analysis) (@data_analysis_ml) in the Russian language segment is an active participant. Currently, the community unites 51 914 subscribers, ranking 2 559 in the Technologies & Applications category and 11 791 in the Russia region.
📊 Audience metrics and dynamics
Since its creation on невідомо, the project has demonstrated rapid growth, gathering an audience of 51 914 subscribers.
According to the latest data from 03 November, 2026, the channel demonstrates stable activity. Although there has been a change in the number of participants by 1 220 over the last 30 days and by -7 over the last 24 hours, overall reach remains high.
- Verification status: Not verified
- Engagement rate (ER): The average audience engagement rate is 9.49%. Within the first 24 hours after publication, content typically collects 8.01% reactions from the total number of subscribers.
- Post reach: On average, each post receives 4 927 views. Within the first day, a publication typically gains 4 158 views.
- Reactions and interaction: The audience actively supports content: the average number of reactions per post is 212.
- Thematic interests: Content is focused on key topics such as llm, контекст, openai, архитектура, deepseek.
📝 Description and content policy
The author describes the resource as a platform for expressing subjective opinions:
“Data science, наука о данных.
@haarrp - админ
РКН: clck.ru/3FmyAp”
Thanks to the high frequency of updates (latest data received on 04 November, 2026), the channel maintains relevance and a high level of publication reach. Analytics show that the audience actively interacts with content, making it an important point of influence in the Technologies & Applications category.
pip install -U mlx-audio
https://huggingface.co/collections/Marvis-AI/marvis-tts-250m-v01-68adf13f5f59206e3910502a
@data_analysis_mlsmart-turn v2 — управление диалогами
— MLX Whisper (large-v3-turbo-q4) — распознавание речи
— Qwen3-235B-A22B-Instruct-2507-3bit-DWQ — основная LLM
— Kokoro — голосовой движок
🧠 Всё это запускается локально, без интернета. Максимальное потребление памяти — ~110 ГБ.
⏱️ Задержка «голос → голос» — примерно 950 мс, и её ещё можно уменьшить на ~100 мс.
💡 Для первых экспериментов — впечатляющий результат. Голосовой AI без облаков уже рядом.
🟢 Github
@data_analysis_ml0 = прямой ответ
- default = без ограничений
- поддержка CoT и саморефлексии
---
📊 Результаты (Seed-OSS-36B-Base)
- MMLU-Pro: 65.1 / 60.4
- MMLU: 84.9 / 84.8
- TriviaQA: 82.1 / 81.9
- GPQA-D: 31.7 / 35.2
- BBH: 87.7 / 87.2
- GSM8K: 90.8 / 90.3
- MATH: 81.7 (SOTA) / 61.3
- MBPP: 80.6 / 74.6
- HumanEval: 76.8 / 75.6
📊 Результаты (Seed-OSS-36B-Instruct)
- MMLU-Pro: 82.7 | MMLU: 87.4
- GPQA-D: 71.4 | SuperGPQA: 55.7
- AIME24: 91.7 (SOTA) | AIME25: 84.7 | BeyondAIME: 65
- ArcAGI V2: 40.6 | KORBench: 70.6
- LiveCodeBench v6: 67.4 (SOTA)
- IFEval: 85.8
- TAU1-Retail: 70.4 (SOTA) | TAU1-Airline: 46
- SWE-Bench Verified: 56 (SOTA) | Multi-SWE-Bench: 17
- MMMLU: 78.4 | RULER (128K): 94.6 (SOTA) | AIR-Bench: 75.6
⚡ Инференс
- Поддержка Transformers и vLLM (≥0.10.0)
- FlashAttention2
- Квантизация 4/8-бит
📌 Итог: ByteDance выкатывает мощный опенсорс-стек для reasoning и агентных задач. Seed-OSS-36B-Instruct бьёт SOTA на множестве бенчмарков — от MATH и SWE-Bench до RULER-128K.
🟢GitHub
🟢Hugging Face