01decide-v1
decision
routing, scoring, classifying. any question with a fixed set of answers, answered with a probability for every option.
small modelsthat decide.
togi is a research lab building small, single-job models for indian languages. not chatbots. models that take a message in hindi, tamil, telugu or hinglish and return one decision, with a number you can trust.
01
decisions, not generations.
a typed question in, one calibrated answer out. no prompt, no sampling, no chain of thought. one forward pass.
02
one model, one job.
the way desert ant builds for on-device audio and vision, we build for indic text. each model does one thing and publishes how well.
03
built for how india types.
lowercase, no punctuation, chat spellings, roman script for hindi and tamil, script switching mid-line. the tokenizer is built for it first.
04
we publish the test we lose too.
every model card carries the benchmark where a competitor beat us. if there is none yet, we say that instead.
togi
to sharpen is to remove everything that is not the edge.
01 · research
three things we work on.
every area ends in a shipped model with a public number. the areas are the lab; the models are the output.
02tongue
language
which language and which script to answer in. the script penalty: what every model loses when devanagari becomes roman hindi.
03toxic · redact
safety
abuse, threats and spam scored rather than generated. phone numbers, upi ids and addresses found and nothing else.
02 · the catalogue
one model, one job.
each model answers a typed question and returns a number you can act on. open weights where the status says so.
4B torso · 36M trained · 171 ms median
language · 9 languagestongue
which language and script to answer in. nine languages, dialect aware.
safety · score 1–5toxic
abuse, threat and spam, scored one to five, never generated.
safety · spanredact
find the phone, the upi id, the address. nothing else.
03 · first result
english-first decision models fall apart on tamil.
753 real-style support messages, four language buckets, nine models scored. this is what decide-v1 was built to pass, and where it still loses.
decide-v1 · live · whatsapp
refund eppo varum 10 days aachu
what is this customer asking for? pick one of eight. score urgency one to five. choose the reply language.
172 ms · one forward pass
intent · 8 options
- order_status0.00
- payment_issue0.00
- refund1.00
- delivery0.00
- account_login0.00
- complaint0.00
- product_question0.00
- other0.00
urgency3 / 5
reply intamil
- intent accuracy
- 0.98
- +0.12 vs pplx-decider-flash 0.86
- tamil intent
- 0.98
- +0.17 vs clef 27B 0.81
- urgency, exact match
- 0.80
- +0.52 vs pplx-decider-flash 0.28
- massive indic, 18-way
- 0.933
- +0.056 vs yama 0.877
togi support benchmark · 753 messages · 8 intents · 4 language buckets
| model | intent | tamil intent | urgency exact | hard cases /24 |
|---|---|---|---|---|
| decide-v1 | 0.98 | 0.98 | 0.80 | 20 |
| clef 27B | 0.88 | 0.81 | 0.34 | 0 |
| pplx-decider-flash | 0.86 | 0.86 | 0.28 | 0 |
| yama | 0.78 | 0.84 | 0.21 | 6 |
- 0.848 indiccopa, cause and effect · pplx 0.737
- 0.842 indiccopa, tamil · pplx 0.622
- 98.8% answers kept after shuffling options · pplx 91.2%
we publish the test we lose too. on one unseen hinglish retail set, pplx-decider-flash scores 0.955 to our 0.940. full tables at togilabs.ai/research.
04 · writing
what we have written down.
- oct 2026decide-v1: a decision model for how india typesone forward pass, calibrated probabilities, built from the tokenizer up for indic scripts and romanised chat.
- oct 2026the togi support benchmark: 753 messages, twelve modelsrouting hindi is close to solved. routing tamil and telugu is not. nobody grades urgency.
- oct 2026the script penaltythe same situation written in devanagari, roman hindi and tamil. what each model loses when the script changes.
05 · about the lab
a small lab, sharpening one edge.
togi means to sharpen. we do not try to make a model that does everything. we take one decision that indian products make a million times a day and make the model that gets it right, in the language the customer actually typed, and we publish the numbers.
