Rohan Jha

474 posts

Rohan Jha banner
Rohan Jha

Rohan Jha

@Robro612

CS PhD Student @jhuclsp Previously: Intern @JinaAI_, MS CS @UTAustin, BS AI @carnegiemellon Interested in Information Retrieval and NLP

Baltimore, MD Katılım Haziran 2015
460 Takip Edilen334 Takipçiler
Rohan Jha
Rohan Jha@Robro612·
@barrowjoseph @antoine_chaffin @LightOnIO I've been wondering the same thing - what's the tradeoff, both in terms of efficiency and class of problems you can solve depending on what tools/affordances (e.g. repl, subagents) you give to your orchestrating model. Tools like Colgrep then are a bicycle for the agent
English
1
0
2
25
Joe Barrow
Joe Barrow@barrowjoseph·
@antoine_chaffin @LightOnIO It feels like there’s a fork in the road between “simple tool, many calls” (DCI) and “expressive tool, few calls” (Agent-ModernColBERT). I can’t help but feel like the latter wins in most cases, especially given the ~quadratic increase in price from many calls.
English
2
0
1
68
Rohan Jha retweetledi
Antoine Chaffin
Antoine Chaffin@antoine_chaffin·
Reason-ModernColBERT nearly solved BrowseComp-Plus, smashing SOTA and outperforming models models 54× bigger Not bad for a 1 year old model not optimized for deep research What if we actually tried? Introducing Agent-ModernColBERT: adding another 10% on top with a 5 min training
Antoine Chaffin tweet media
English
11
44
224
38.9K
Rohan Jha
Rohan Jha@Robro612·
@AmelieTabatta > if any stage of your training mix touched MS MARCO, the leaderboard delta is noise. Wasn't the point of BEIR to measure MS MARCO-trained models' ZS performance on OOD (non-MARCO) data? cc @beirmug
English
1
0
2
113
Amélie Chatelain
Amélie Chatelain@AmelieTabatta·
Half of 'our embedder SOTA on BEIR' claims are measuring contamination, not retrieval quality. If any stage of your training mix touched MS MARCO, the leaderboard delta is noise. Hold-out benchmarks aren't optional anymore, run your own evals!
English
3
2
26
2.5K
Rohan Jha
Rohan Jha@Robro612·
@mixedbreadai Loved the examples! Very instructive of what well-trained instructed rerankers can unlock.
English
0
0
4
547
Mixedbread
Mixedbread@mixedbreadai·
Introducing mxbai-rerank-v3-listwise: reranking that goes beyond binary relevance. It reads the whole candidate set, resolves conflicts, and ranks by directives like recency, source priority, and multi-step rules. +11% NDCG@10 on average across multiple domains, modalities, and languages in runs with Wholembed v3. Available today in preview in Mixedbread.
Mixedbread tweet media
English
5
18
137
24.2K
Rohan Jha
Rohan Jha@Robro612·
@n0riskn0r3ward I think of @macavaney & @rathee_mandeep 's adaptive retrieval line of work as falling under PRF, and are really exciting approaches I hope we'll see adopted more. Papers: GAR, QUAM, and Breaking the Lens of the Telescope
English
0
0
2
107
search founder
search founder@n0riskn0r3ward·
Getting re-interested in pseudo relevance feedback papers - if you have any favorites pls share them!
English
2
0
2
237
Rohan Jha retweetledi
Antoine Chaffin
Antoine Chaffin@antoine_chaffin·
XTR allows to perform multi-vector retrieval faster But there is not much models and tooling around it, hindering its adoption @Robro612 did a very interesting replication study and we took the opportunity to merge XTR into PyLate, alongside the awesome XTR-WARP of @hugemensa
Antoine Chaffin tweet media
Rohan Jha@Robro612

New 📄: we replicate XTR, a multi-vector retrieval method that makes ColBERT faster by avoiding its expensive step of gathering full document embeddings XTR is not a free lunch over ColBERT, but its training objective is useful for modern efficient engines like PLAID and WARP 👇🏼

English
7
10
47
5.4K
Antoine Chaffin
Antoine Chaffin@antoine_chaffin·
@hugemensa @Robro612 @raphaelsrty > albeit somewhat nerfed If you refer to the parameters, power users can set their own trade offs If you refer to all of the other amazing features, I am very much looking forward to adding them all!
English
2
0
2
169
Pau
Pau@hugemensa·
Thanks to the amazing work from @Robro612 , @raphaelsrty and @antoine_chaffin v2 of xtr-warp-rs is available in PyLate! (albeit somewhat nerfed) XTR models can now be trained from PyLate as well, a great release that gets mainstream adoption of multi vector closer!
Antoine Chaffin@antoine_chaffin

XTR allows to perform multi-vector retrieval faster But there is not much models and tooling around it, hindering its adoption @Robro612 did a very interesting replication study and we took the opportunity to merge XTR into PyLate, alongside the awesome XTR-WARP of @hugemensa

English
1
1
6
728
Rohan Jha retweetledi
Raphaël Sourty
Raphaël Sourty@raphaelsrty·
XTR training and WARP indexes are available in PyLate 1.5.0 Credit to @Robro612 for the XTR integration and @hugemensa for the WARP index ☺️ The WARP index can run on GPU and will shine when models are trained with XTR procedure
Rohan Jha@Robro612

New 📄: we replicate XTR, a multi-vector retrieval method that makes ColBERT faster by avoiding its expensive step of gathering full document embeddings XTR is not a free lunch over ColBERT, but its training objective is useful for modern efficient engines like PLAID and WARP 👇🏼

English
1
6
28
3.4K
Rohan Jha
Rohan Jha@Robro612·
Fortunately, thanks to the diligent work of open-source GOATs @antoine_chaffin @raphaelsrty @hugemensa, XTR and WARP are now integrated into PyLate. Trying it yourself should be as simple as swapping the score function during training and the index at inference.
Rohan Jha tweet media
English
2
0
7
169
Rohan Jha
Rohan Jha@Robro612·
New 📄: we replicate XTR, a multi-vector retrieval method that makes ColBERT faster by avoiding its expensive step of gathering full document embeddings XTR is not a free lunch over ColBERT, but its training objective is useful for modern efficient engines like PLAID and WARP 👇🏼
English
1
25
63
11.4K

Keşfet

@barrowjoseph @antoine_chaffin @LightOnIO @AmelieTabatta @beirmug @mixedbreadai @n0riskn0r3ward @macavaney