Saka

Drifting through eldritch seas, where dark matter dreams and reality glitches.

[Saka @SakaSurya_]

@theviclaw 🤣🤣🤣🤣

[Saka @SakaSurya_]

@lochan_twt Unless fable gonna stop in middle due to guardrails

Oops you should use opus

[Saka @SakaSurya_]

@0xRemakw Its helpful to lawyers, moreover lawyers income gonna rocket 🚀

[Saka @SakaSurya_]

@rohanpaul_ai How fascinating legal tech companies giving to claude ,so that claude can eat them later

[Saka @SakaSurya_]

@iamkylebalmer Legal is 3rd after software, gonna have huge impact by AI

[Saka @SakaSurya_]

@willchen500 Sovereign application stack deeply rooted might be helpful , each model has its own capabilities

[Saka @SakaSurya_]

https://t.co/1Ffg3LsVqs

[Saka @SakaSurya_]

We are releasing counsel 27b legal and general model reduce thinking and increase capabilities of tool call, etc

Achieved 4.3x speed on single gpu ( around 220 tokens/s) and 6x on 2 gpus (350 token/s ) and in few use cases upto 3 to 4k token/s

On amendments or tiny changes

4 times speed with less than half of the price

https://t.co/BKtje1R0aF

[Saka @SakaSurya_]

What does cohere even thinking, they can come out of internet after blaming qwen without evidence.

[Saka @SakaSurya_]

https://t.co/EmsPhPuSfu

[Saka @SakaSurya_]

We tested whether models can improve predictions during inference and what context should go into weights or context, tested on attention , hybrid attention, diffusion architectures.

5 to 8 percent of software adaptation in attention style,

8 to 20 percent in diffusion style and less than 1 percent in hybrid attention.

What we found is

weight updates are useful for adapting how models process information( docs ) and Distillation for reasoning improvement

[Saka @SakaSurya_]

@timshi_ai Seriously, its 1/100k as they mentioned and no logs, traces ,code ? And pure FUD, why not come clean

[Saka @SakaSurya_]

@1vnzh Dude, why dont you release traces, logs and code????

[Saka @SakaSurya_]

@aidangomez Why don't you release traces??

[Saka @SakaSurya_]

@cohere You should release the traces , so community can dig deeper

[Saka @SakaSurya_]

AI gonna make legal services affordable

[Saka @SakaSurya_]

we open-sourcedCounsel-2B, a compact legal language model. It can work in both cpu or gpu has capabilities of

Thinking mode, function calling and upto 128K context and a reasoning core for legal agent workflows. It also ships with Gates, gates helped the model double its score on Harvey LAB legal agent tasks from 8.9% → 18.1% of criteria passed.

Apache-2.0

[model link](https://t.co/2Vx3zFoPf3)https://t.co/Gx3z684M2J

[Saka @SakaSurya_]

💼 Bharat Economy

Model (address resolution)
- BharatAddr: resolves free-text Indian addresses to the correct village 95.7% of the time, vs 86.4% for a strong baseline.
https://t.co/sM0LinzH7Y

Model (industry-code classifier)
- NIC-Coder-0.6B: assigns a business description its NIC industry code. Its top-1 accuracy is 16.4% vs 7.2% for TF-IDF.
https://t.co/uUdtxrRVVP

Datasets
- BharatAddr-Gazetteer: the place-name gazetteer behind BharatAddr.
https://t.co/YyW77SWjPq
- Bharat-Workforce-Atlas: district counts of eShram-registered unorganised workers by occupation, gender and age.
https://t.co/FLExLXgrya
- Bharat-Skills-Atlas: district profiles of India's unorganised workforce.
https://t.co/RXQYADPhoL

💧 Bharat Water

Model (time-series forecaster)
- BharatWater-Chronos: groundwater and reservoir forecasting. It has lower error than Chronos-2 at every horizon: 7 days 0.263 vs 0.268, 30 days 0.293 vs 0.305, 90 days 0.333 vs 0.374.
https://t.co/NfPKZRdoSO

Dataset
- BharatWater-Data: https://t.co/Sgur6UWa1y

Benchmark
- BharatWater-Bench: https://t.co/NtWSGr0ODO

Leaderboard
- https://t.co/Ar8ApLfXMZ

⚡ Bharat Energy

Model (time-series forecaster)
- BharatEnergy-Chronos: power-demand forecasting. It has lower error than Chronos-2 at every horizon: 1 day 0.090 vs 0.092, 7 days 0.138 vs 0.143, 30 days 0.176 vs 0.186.
https://t.co/4yoqNqPntH

Dataset
- BharatEnergy-Data: https://t.co/YBBgNesgc7

Benchmark
- BharatEnergy-Bench: https://t.co/Y1QOCkFvpi

Leaderboard
- https://t.co/kqMpkDNV4f

---

If you work on public policy, agriculture, public health, civic tech or Indian-language AI, we'd love to hear what you build with it.

Browse every release:

https://t.co/cyE8468RRm

#OpenData #India #AI #MachineLearning #HuggingFace #OpenSource #DigitalIndia #GovTech #DataForGood #AgriTech #PublicHealth

[Saka @SakaSurya_]

We open-sourced 13 models, 20 datasets, 8 benchmarks and 6 live apps built on India's open government data

India publishes one of the largest open-data collections in the world, through https://t.co/nO8H1fM6MR, NITI Aayog's NDAP and dozens of ministry portals. There are hundreds of thousands of tables on health, agriculture, water, energy, employment and district development.

But most of that data is hard to use. It is split across separate CSVs with inconsistent formats, units and district names, and there are no benchmarks or models built for it.

Over the past few days we turned it into a full, open stack across eight domains. To be clear about what is in it, since not all of it is AI models:

- 13 models. Of these, 5 are time-series forecasters, 3 are fine-tuned LLMs (all Qwen3-4B), 2 are embedding models, 1 is a tabular foundation model, 1 is an address-resolution model and 1 is an industry-code classifier.
- 20 datasets, cleaned and typed.
- 8 benchmarks
- 6 live apps: 5 open leaderboards and 1 MCP server.

Everything is on Hugging Face: https://t.co/cyE8468RRm

Below is what we released in each domain, grouped by what each thing actually is.

---

🏥 Bharat Health

Health data from India's Health Management Information System (HMIS) and disease surveillance, turned into forecasting tools and audit data.

Model (time-series forecaster)
- HMIS-Chronos: Chronos-2 fine-tuned to forecast Indian health indicators month by month.
https://t.co/5iOqilEMPf

Datasets
- IndiaDiseaseWatch: district × month counts of 28 disease, morbidity and context items reported by public and private health facilities.
https://t.co/WSDfm2eXi2
- Bharat-Health-Service-Atlas: how well each district delivers basic health services, year by year.
https://t.co/ywHFGUINwn
- HMIS-Audit: cell-level data-quality findings for HMIS sub-district monthly reports.
https://t.co/fsUqlJYyfN
- Health-StatsQA: 1,000 health-statistics questions about Indian districts, each in English and Hindi.
https://t.co/InGH9Hb3AC

Benchmark
- HMIS-Bench: a probabilistic forecasting benchmark for Indian health data.
https://t.co/IGUJ2dCQHT

Leaderboard
- https://t.co/du4eiCtvdz

🌾 Bharat Agriculture

Built from nearly 5 million Kisan Call Centre conversations, mandi prices and crop statistics.

Model (fine-tuned LLM)
- KCC-Advisor-4B: Qwen3-4B fine-tuned on Kisan Call Centre question-and-answer pairs. It passes our safety judge 94.4% of the time vs 88.7% for the base model, and has a built-in crisis guard. Research use only; not agricultural advice.
https://t.co/faZDPRmxcc

Model (embedding)
- KCC-Embed-0.6B: an embedding model for Indian farmer questions.
https://t.co/PsOTjnp3Lu

Model (time-series forecaster)
- Mandi-Chronos: forecasts daily wholesale mandi prices.
https://t.co/BBrpGwfnI9

Datasets
- KCC-Pairs: 4,996,402 farmer-question to advisor-answer pairs.
https://t.co/DRA8ohRwrl
- KCC-Pulse: weekly Kisan Call Centre call counts per district and theme (pest, disease, weed, nutrient).
https://t.co/5l5MKN8x1q
- PMKISAN-Reach-Atlas: who PM-KISAN reaches, instalment by instalment, from village to state, for instalments 1 to 22, including women's share.
https://t.co/MSPyVFH4IR

Benchmarks
- MandiBench: a probabilistic benchmark for daily mandi prices, with 12,863 test series.
https://t.co/Bo5e8QJT1L
- CropBench: forecasting district crop yield and production.
https://t.co/El0bP1NuDF

Leaderboard
- https://t.co/Oz9QUXO4rq

[Saka @SakaSurya_]

📊 Bharat Tables

The core of the release: India's open government data as clean tables that you can search, query and talk to.

Models (fine-tuned LLMs)
- BharatSQL-4B: turns a plain-English question into SQL over Indian government tables. It scores 91.3% exact match, ahead of the 27B model it learned from (83.9%) and base Qwen3-4B (74.6%).
https://t.co/YubxUACMdz
- BharatTableScribe-4B: turns a messy raw government table into a clean, standard table and writes the cleaning code to reproduce it. It scores F1 0.974 on NDAP tables vs 0.535 for a rule-based cleaner.
https://t.co/TPnfW3VuqE

Model (embedding)
- IndiaTables-Embed-0.6B: finds the right government table from a question in English, Hindi or Hinglish. It scores nDCG@10 0.839 vs 0.683 for the base model and 0.490 for BM25 keyword search.
https://t.co/HDfsfddfzc

Datasets
- IndiaTables: 266,707 cleaned, typed tables from https://t.co/nO8H1fM6MR and NDAP.
https://t.co/f58JgT5QvO
- India-Data-Catalog: one clean, searchable metadata row for every https://t.co/nO8H1fM6MR resource.
https://t.co/KisE2BpcAS
- BharatSQL-Data: 262,538 execution-checked text-to-SQL pairs.
https://t.co/uOoBjAlOsy
- BharatTableScribe-Data: raw tables paired with clean versions and executable cleaning code.
https://t.co/3bGlr5Ry4U

Benchmarks
- IndiaStatsQA: 20,000 numeric questions about Indian state, district and village statistics.
https://t.co/fWQdYEtBrC
- IndiaTables-Bench: five tasks over Indian government tables.
https://t.co/6Q0WzDEOLK

Leaderboard
- https://t.co/BauPe59zRg

MCP server
- OpenGovtData-MCP: an MCP server and HTTP API over IndiaTables. Plug India's open government data straight into Claude, ChatGPT or any agent.
https://t.co/wjjqcgNTXZ

📈 Bharat Forecasting

Model (time-series forecaster)
- BharatTS-Large: a probabilistic forecaster for Indian government time series, from mandi prices to health indicators. On our held-out set its error (WQL) is 0.171 vs 0.207 for Chronos-2.
https://t.co/x6qN7DlAqF

Model (tabular foundation model)
- BharatTab-FM: a tabular foundation model for Indian administrative data. It predicts on a new table straight away, with no training per dataset.
https://t.co/3pJdgkFStN

Dataset
- BharatTS-Corpus: 1,799,285 time series (76.3 million observed points) from 8,461 government tables.
https://t.co/BJYWP9H6t8

Benchmark
- BharatTab: 17 district-level prediction tasks: crop yields, health, PM-KISAN, schools and MSMEs.
https://t.co/CGsZOoqWSY

🏛️ Bharat Governance

No models here, just data.

Datasets
- BharatDistrictPanel: one district × year panel that joins sixteen official sources on current district boundaries.
https://t.co/dogjz9uPri
- India-District-Briefs: short English and Hindi briefs for 794 districts.
https://t.co/moUAnrxhfy
- BharatData-Audit: an arithmetic and consistency audit of 266,707 https://t.co/nO8H1fM6MR tables. It flags where parts don't add up to the total, where male + female ≠ total, and where percentages are out of range.
https://t.co/PQZLOTT8mG