Bad Ideas

• Fine-tuning instruct models. The core idea for further fine tuning of instruct models, did not extend beyond 1000 examples for newer models which are much much better for inference. The core mistake as highlighted in the BioMixtral paper was to look at continual pretraining followed by full fine tune to stop the problem of chaotic learning or catastropic forgetting instead of touching the chat model.

https://arxiv.org/pdf/2402.10373.pdf

What Worked?

• Rerankers: It seems that a lot of dimensions in the vector embeddings of our BGE-m3 fine-tune (1024) and even the latest openai embeddings, cluster all our text quite closely, making it hard to pick top n clearly. This is solved by our usage of cross-encoder models to rerank retrieved chunks vs user questions. This increases speed of inference ofc

💡 We have even tried using cross-encoder models for grading multiple LLM config answers as an alternative to GPT4 autograding. This is detailed here: github.com/never2average/frugalgrader and will be updated with our fine-tune decoder and mongoDB integrations soon. Saves us time on UAT and

• Vision Pipelines: Slides are designed to be viewed and not computed on, therefore a language model will not capture the whole structure. However, 0 shot text extract from vision models is not sufficient to capture context which relies on location on interpreting images. However, when coupled with OCR output in context, even OSS models like LLaVa and SOTA GPT-4V give great structured representations of the page which are useful for further use.

W/O Vision

 Disclaimer: The information contained in this section has been obtained from third parties. These figures maynot necessarily have been independently verified. We do not assume responsibility for the accuracy orcompleteness of such information.WATERBRIDGE VENTURES
Fund VintageYearFund Size IRR (INR) (Netof expensesand pre carry)DPI (INR)TVPI(INR)MOIC(INR)WBV I 2017INR 140.8 cr 21%0.7x2.4x2.8xWBV II2020INR 1000 cr28%-1.7x2.1xWBV Alpha(Continuity)2022INR 241.5 cr NMFNMFNMFNMF
VC FUNDEARLY STAGE  CAT II AIF
Started and operated by experienced investment professionals witha combined 40 years of investing experience for all 3 partnersThesis-driven investing in companies with entry point at seed andpre-series A rounds with a tech angle where the market segment islargeTheir strategy is to build a curated portfolio wherein they’d enterearly, maximise returns across sectors and going deep in thewinnersThe first two funds are both in the top performance quartileindicating consistent performanceABOUTMANISH KHETERPAL  Co-founder Managing Partner  Stanford, IIT DelhiPE + VC with global fundsfor 19 years, worked withRho, Providence, andActis ASHISH JAIN Co-founder Partner DU, CA(ICAI)Growth, PE, and FundOperations with Actis, LokCapital, Xander andEscortsANJALI SOSALE PartnerBangalore University,CA(ICAI)Corporate VC withMyntra, CorporateFinance and IB, StartupConsulting for EYPAST TRACK RECORD (AS OF DEC’23)  WATERBRIDGE VENTURES HIGHLIGHTS
KEY PORTFOLIO COMPANIES33~INR 1382 cr34Delhi, BangaloreAgnosticEarly Stage20174
TOTAL FUNDSTOTAL FUNDS RAISEDACTIVE SINCE
FULL EXITSPORTFOLIO COMPANIES
LOCATIONSECTORSSTAGE

WATERBRIDGE VENTURES
VC FUND: EARLY STAGE | CAT II AIF
ABOUT
Started and operated by experienced investment professionals with a combined 40 years of investing experience for all 3 partners
Thesis-driven investing in companies with entry point at seed and pre-series A rounds with a tech angle where the market segment is large
Their strategy is to build a curated portfolio wherein they'd enter early, maximise returns across sectors and going deep in the winners
The first two funds are both in the top performance quartile indicating consistent performance
| Fund     | Vantage Year | Fund Size    | IRR (INR) (Net of expenses and pre carry) | DPI (INR) | TVPI (INR) | MOIC (INR) |
|----------|--------------|--------------|------------------------------------------|-----------|------------|------------|
| WBV I    | 2017         | INR 140.8 cr | 21%                                     | 0.7x      | 2.4x       | 2.8x       |
| WBV II   | 2020         | INR 1000 cr  | 28%                                     | -         | 1.7x       | 2.1x       |
| WBV Alpha| 2022         | INR 241.5 cr | NMF                                     | NMF       | NMF        | NMF        |
 
MANISH KHETERPAL
Co-founder
Managing Partner
Stanford, IIT Delhi
PE + VC with global funds for 19 years, worked with Rho, Providence, and Actis
ASHISH JAIN
Co-founder Partner
DU, CA(ICAI)
Growth, PE, and Fund Operations with Actis, Lok Capital, Xander and Escorts
ANJALI SOSALE
Partner
Bangalore University, CA(ICAI)
Corporate VC with Myntra, Corporate Finance and IB, Startup Consulting for EY
WATERBRIDGE VENTURES HIGHLIGHTS
~INR 1382 cr TOTAL FUNDS RAISED
3 TOTAL FUNDS
3 FULL EXITS
34 PORTFOLIO COMPANIES
Early Stage STAGE
Agnostic SECTORS
Delhi, Bangalore LOCATION
KEY PORTFOLIO COMPANIES
magicpin
CHALO
atlan
CITYMALL
bijins
pocketpills
 
 

[Preprint soon]

• Logical Separation of Formats for Retrieval: This should have been kind of trivial but wasn’t apparently. The information density across tokens isn’t uniform for text, tables and charts. Specially since we embed questions generated from the content, rather than the raw content itself for multiple reasons, 1 being roughly uniform content size.

What took longer than expected?