Every {}, "field", :, and comma in JSON becomes tokens.
LLMs tokenize JSON poorly, inflating both latency and API cost.
TOON removes the noise and keeps only the meaning.
#ToonStream
LLMs aren’t expensive.
Your JSON is.
Switching from JSON → TOON cuts token usage by:
55% on flat data
38% on arrays
15% on nested objects
Here’s why this matters — and how much you can save. 👇
#LLM#AI#AGENTS#RAG
Meet Toonstream, the library that implementing
TOON (Token-Oriented Object Notation) —
a serialization format built to make structured LLM inputs far more token-efficient than JSON.
! pip install toonstream
#LLM#AiAgents
I have a classification problem, and XGBOOST is the best model.
I have calibrated the model outputs as well and saved both the models.
Say, I am creating a pipeline to predict for future values.
Which model should I use?
Normal or calibrated model ?
@tunguz#DataScience
`plt.subplot_mosaic(...)` is the single-most amazing @matplotlib function I'd never heard of 😍🤓🌍 Can't believe I've used Python for more than a decade and only just discovered it! Subplots will never be the same again 🌟 https://t.co/EBN4ndA6HL
@haltakov evaluation metric is totally depend on the type of data we have.
If it's balanced data, we can use accuracy.
If it's imbalance, we can either use precision or recall depends on the use case.
Python 3.10 is out, every damn programmer is in love with the new structural pattern matching.
But for anyone teaching python out there, the new error messages are what's up. (Before/After)
Ecstatic to announce that I have joined Forbes as a Data Scientist.
I am blessed to have landed in 5 offers (Data Scientist, MLE) from 1000+ applications, 30+ recruiter calls, 20+ technical interviews, 15+ Online assessments & 10+ final rounds.
#DataScience#machinelearning