Proud to announce:
💫 garak - an LLM vulnerability scanner💫
🔎 Check if a model is susceptible to common attacks
🦜 Supports HuggingFace, OpenAI, ggml, Cohere, ...
🔧 >70 probes: prompt injection, false claims, toxicity, encoding evasion, ..
ChatGPT not best at many language tasks. It's outranked by other systems on many NLP benchmarks in current evaluation. For 77.5% of tasks examined, other systems are better than ChatGPT.
Is Cosine-Similarity of Embeddings Really About Similarity?
Netflix cautions against blindly using cosine similarity as a measure of semantic similarity between learned embeddings, as it can yield arbitrary and meaningless results.
📝
ELIZA designer Joseph Weizenbaum observed: “What I had not realized is that extremely short exposures to a relatively simple computer program could induce powerful delusional thinking in quite normal people.”
Efficient NLP methods - an up-to-date survey, to appear in TACL. We cover efficiency wrt:
* Data
* Model design
* Pre-training
* Fine-tuning
* Inference & compression
* Hardware utilization
* Evaluation
* Model selection
This was a blast to co-produce!
How can NLP be more efficient? We can't blindly scale forever. This survey I'm lucky to be part of presents current efficiency research at many stages of the NLP process, and actionable advice to practitioners for making their NLP more efficient.
#nlproc
the replicability crisis has been real for over a decade, but replicability still is nowhere near a publishing requirement in our field. even hugging face are ahead of our publishing venues in this and they're just.. 3? 4? years old (from here )
"Dear Leon, I’m happy to say that the CS department has decided to offer you a position as associate professor" -- What a privilege! Only 6.5 hours til the bubbles can be opened :)
@JeffDean
Submitting insufficiently PR-washed work *for review* is a summary terminable offence at Google, especially if you have a minority background? Got it, thanks for the clarification
We release T5 large language models for Danish, the biggest monolingual models for the language. We had to use a bunch of tricks to get this to work with modest data and hardware; here's how:
#dktech
There's an NLP task, "detoxification", where the goal is to rephrase a statement that uses toxic language into one that doesn't use toxic language. This is a really odd task laden with harms. ⬇️
you can even get strong papers with new results by doing replicability studies - here's an example, - and there are more every year. so why do we accept so many results on trust, with no replication? because there's no incentive to confirm findings
Streamers trying to game location-based content recommendation algorithms by going to a wealthy neighbourhood to stream. For donations. Under a bridge. Sat on the ground.
New paper! 📝
Text data can contain harmful things, but unlike other disciplines, we don't have protocols for handling or presenting harmful material🛢️. This paper with
@hannahrosekirk
@Abebab
@bertievidgen
presents the risks & some solutions. 🥼🥽
What toxic nonsense Google have chosen to autoinclude in their search results. Irresponsible use of AI with Google then taking no ownership over the consequences of their actions when told about them - doubly embarrassing.
@WendaXu2
Oh yeah, I've had this a few times. Officer is correct. I always make sure I have a book, now, to pass the time. LAX is particularly rude, ORD has an even less comfortable waiting area - recommend SFO or SEA for ingress, they tend to be polite and efficient
Love this
* attacker can inject backdoors by issuing very few malicious instructions among thousands of gathered data
* attacker can achieve > 90% success
* instruction attacks show resistance to existing inference-time defense
Any mTurker has control!
Integrating electricity consumption into loss during hyperparameter search finds LLMs that are pareto-efficient for power and benchmarks, doing the same as bigger models for less.
"Hyperparameter Power Impact
in Transformer Language Model Training" -
Debunking the idea that using linear activations means a computational neural net simplifies to an affine function of the input layer. We don't compute with continuous value representations, so the simplification is missing. Quantisation is powerful!
@RealSexyCyborg
@pixelgem
Please understand that changing the question derails the conversation -- a technique frequently used in classic Soviet propaganda -- and deliberately derailing a conversation about gender equity is inescapably sexist. That's what you're doing here.
@adityas129
oh man i used to do this!
* dedicate time to working on your emotional and mental health
* drop every single extra project
* close the computer and turn off the phone every now & then
* give focus *only* to non-work things outside of 9-5
* don't have kids. DO NOT have kids
Women in Denmark experience the most violence out of *all EU states*, including both from their partners and from outside a relationship. Analysis is thorough and addresses reporting biases etc. well:
#twitterhjerne
#dkpol
If GPT3 took 50 petaflop-days to train , w. GPUs at 10^8 flops per watt , so those 1.2E18 flop-hours used 12 GWh to train? E.g. 12 hours of a whole nuclear reactor? At 0.73kg per kWh that's.. 8.8 megatons of CO2?!
#sanitycheck
#nlproc
@ClaudiaStokes
This casts the net inappropriately broad, doesn't it. Met my wife at a university bar while she was a PhD candidate and I was a postdoc, in different departments; now I'm a professor, she's a former grad student, and nothing was wrong.
Modeling the rhetoric of persuasion, with application to propaganda and disinfo detection -- accepted to NLP for Positive Impact at EMNLP, with Amalie Brogaard Pauli and Ira Assent. Really excited about this direction
@SwiftOnSecurity
Right. From our years of work on fake news (e.g.
@PhemeEU
), it become evident that planted stories that "feel" right and sound edgy are those consumed most readily
Seems the majority of
#COLING2018
papers are presented with code/data, on github or even "pip install <x>". This is the first
#nlproc
conference I've seen where sharing is the overwhelming norm; not sharing is now the unusual, minority behavior.
#openscience
Clever Hans: the horse we thought could do arithmetic, but really was relying on other signals. Really enjoyed this blog post on NLP, clever hands and what to do instead of leaderboards.
#nlproc
@bryancsk
there is ice *everywhere* here and I don't know why. some random floor of a hotel? you can get rooms, stairs, elevators, a fire hose, and ICE. gas station? we have gas, diesel, and ICE
#respectallcultures
Danish Gigaword v1.0, the largest collection of digital text in Danish (> 1 000 000 000 words), is now released and freely available:
Now we have the resources to do serious AI in Danish; I'm so excited to see the next step.
#nlproc
#dktech
The Broad Twitter Corpus v1.0 is released! - 150K tokens of tweets spread over time, place, and social use. This corpus is also a
@github
repo, so please send pull requests and create issues. Paper:
#nlproc
Our paper on data for abusive language detection is here. Highlights:
- Best practices for making abusive language data
- Advice for machine learning practitioners
- Emerging research problems
- Comprehensive corralling of the state-of-the-art
#nlproc
Language detection is hard, especially between similar languages. For example, the best tools still mix up Danish and Norwegian. Here's a new challenge for distinguishing between six similar Nordic languages, with dataset and baseline. Enjoy!
#nlproc
Funeral today. That leaves us with 0/8 grandparents, 1/4 parents (disabled and abroad), two us, and 1/n children. People with families that help out, you are fortunate!
To the fragile men arguing against the findings' validity: What's your end goal here? Who are you trying to silence, and what good would that serve? Sometimes the best first step to improving a problem is just to listen to and absorb the facts. Read the report.
@ChrisMurphyCT
* It was explicitly built into the model.
* The people building it programmed it to use data including complicated chemistry.
* It has zero agency and didn't decide a single thing for itself.
acl: just saw that our author response lifted a 2 -> 3 and a 2.5 -> 3.5, and then the paper was accepted to the main conf (scores out of five). this pleases me.
Time after time, datasets without documentation turn out to be stuffed with toxic, unjust, and illegal data. The models trained on them propagate the harms. This time, it's the Stable Diffusion data that amazingly has no docs. -- Really thought this lesson had been learned by now
The Impossible Safety of Large AI Models
thank you. even with my AI Ethics hat off and my CS/infosec hat on, this seemed super likely.
"this paper systematizes our knowledge about the fundamental impossibility of building arbitrarily accurate and secure machine learning models"
In this paper,
@L_badikho
@Sadegh_Farhad
@rlt_s
other colleagues and I, who have written many of the AI security papers published at top venues, reviewed the AI security literature and conclude that securing large models is mathematically
#impossible
.
Back to full time industry work today for the first time in 19 (!) years. I don't have a "quit lit" about academia, it was and is a weird, special place where people do wonderful things. Besides, I'll keep my uni affiliation! OK, see you out there in the usual places :)
@yoavgo
reward dataset papers, so people bother producing them. i've made a ton of smaller datasets to enable research on new problems and framings - it's way too much effort, you won't get a good paper, and it'll work against one's career. this is a cultural problem.
nothing made me more grateful for academia than moving to industry. the amount of academic research we consume, implement, and rely on to grow is astounding. i've never pored over such Volumes of papers before. thank you, profs postdocs grad students and undergrads, you rock! ❤️
New paper: "The Danish Gigaword Corpus" - a clean, open, billion-word corpus of Danish stretching across domains, modalities, time, dialect, ... To be presented at
#nodalida
w/
@ManuelCiosici
#nlproc
"Efficient Transformers with Dynamic Token Pooling"
A family of language models that pool variable-size segments of tokens in intermediate layers to enhance the efficiency & performance
We looked at misogyny online. Getting the process right is important. We present: a detailed annotation methodology & codebook; a comprehensive taxonomy of labels for annotating misogyny; and a high-quality dataset of annotated posts; in this ACL paper.
Big text datasets should be cleaned up a bit, for e.g. hate and abuse, before distribution. People who collate and create datasets are responsible for the selection of content in those datasets. So, they should do the cleaning.
Super happy to announce ACL SIGSEC, a Special Interest Group on NLP & LLM Security
* Everyone is welcome
* Regular invited talks (see the schedule!)
* Explicitly connecting industry and academia
* Annual open research workshop
Join us!
1. LLMs are very useful
2. LLMs present concrete dangers & risks
-> These two claims don't conflict
People who state one of these claims don't actually refute the other, despite the assumptions around in the discourse here (this IS twitter, right?). Most people accept both.
Online hate pushes people out of conversations and polarises online societies. We took a scientific approach using machine learning to analyse online hate against women, the largest target group.
Video ->
#hatespeech
#misogyny