I've scored an exciting new opportunity - helping customers avoid disaster when using llm based tooling using systems architecture.
-
I've scored an exciting new opportunity - helping customers avoid disaster when using llm based tooling using systems architecture. I'm doing a presentation and a 2 day workshop.
As part and parcel for this offering, I'm doing a massive shitload of reading up on details and particulars of the entire ecosystem, so I can arm attendees by polishing their brainmeats.
today I found this: https://sites.google.com/view/llmwritingdistortion/home
-
I've scored an exciting new opportunity - helping customers avoid disaster when using llm based tooling using systems architecture. I'm doing a presentation and a 2 day workshop.
As part and parcel for this offering, I'm doing a massive shitload of reading up on details and particulars of the entire ecosystem, so I can arm attendees by polishing their brainmeats.
today I found this: https://sites.google.com/view/llmwritingdistortion/home
if you dont even read it, and you skip straight to the bottom, the little sig there? with the curly braces?
thats how you format input data for unsloth/lora style training.
google writes a paper about how llms are introducing drift to our language and its semantics, and they sign it using a tag intended to be imbibed by something like unsloth.
so this is like.. the smoker talking about how smoking fucks people up, then lighting up a cigarette at the end of the monologue?
-
if you dont even read it, and you skip straight to the bottom, the little sig there? with the curly braces?
thats how you format input data for unsloth/lora style training.
google writes a paper about how llms are introducing drift to our language and its semantics, and they sign it using a tag intended to be imbibed by something like unsloth.
so this is like.. the smoker talking about how smoking fucks people up, then lighting up a cigarette at the end of the monologue?
i'd argue the contents of the paper are an important read, but like.. signing it like that at the end is a huge fucking tell about the bias here
-
i'd argue the contents of the paper are an important read, but like.. signing it like that at the end is a huge fucking tell about the bias here
@Viss or the authors just showing off.. “Watch this guys, I’m going to sign off ironically”
-
i'd argue the contents of the paper are an important read, but like.. signing it like that at the end is a huge fucking tell about the bias here
@Viss that aside, would be interested in your talk
-
if you dont even read it, and you skip straight to the bottom, the little sig there? with the curly braces?
thats how you format input data for unsloth/lora style training.
google writes a paper about how llms are introducing drift to our language and its semantics, and they sign it using a tag intended to be imbibed by something like unsloth.
so this is like.. the smoker talking about how smoking fucks people up, then lighting up a cigarette at the end of the monologue?
@Viss um, that's a ~40-year-old reference standard format called BibTex https://www.bibtex.org/Format/
-
@Viss um, that's a ~40-year-old reference standard format called BibTex https://www.bibtex.org/Format/
@hrbrmstr huh. i guess that format is probably where unsloth is taking its cues - https://unsloth.ai/docs/get-started/fine-tuning-llms-guide/datasets-guide
-
@Viss that aside, would be interested in your talk
@Cali at present its being custom tailored for this client, but a lot of the same raw materials are going into my 'claude is your insider threat now' talk - which sadly securityfest passed on. maybe labscon or sec-t pick it up?
-
@hrbrmstr huh. i guess that format is probably where unsloth is taking its cues - https://unsloth.ai/docs/get-started/fine-tuning-llms-guide/datasets-guide
@Viss gotta pick some delimeters (delimieters gonna delimit?
)FWIW the {} syntax in Unsloth's to_sharegpt is Python's standard str.format() / f-string placeholder convention. It's likely just a path of least resistance: "Hello {name}".format(name=value) is idiomatic Python, and using the same delimiter means column substitution reads naturally to any Python developer and requires zero custom parser logic — just .format(**row) or str.format_map() against a dict of column names.
-
@Viss gotta pick some delimeters (delimieters gonna delimit?
)FWIW the {} syntax in Unsloth's to_sharegpt is Python's standard str.format() / f-string placeholder convention. It's likely just a path of least resistance: "Hello {name}".format(name=value) is idiomatic Python, and using the same delimiter means column substitution reads naturally to any Python developer and requires zero custom parser logic — just .format(**row) or str.format_map() against a dict of column names.
@hrbrmstr admittedly, im not a super expert with this stuff yet, so visually, the immediate reaction was like "wtf, are they trying to be funny or ironic here?"
-
R relay@relay.infosec.exchange shared this topic