:trebuchet: Kale :trebuchet:
@darkestkale.mastodon.social.ap.brid.gy
It'll all be fine.
[bridged from https://mastodon.social/@DarkestKale on the fediverse by https://fed.brid.gy/ ]
[bridged from https://mastodon.social/@DarkestKale on the fediverse by https://fed.brid.gy/ ]
Alright, let's fucking do some epochs, see what's crackalackin
December 3, 2025 at 2:28 PM
Alright, let's fucking do some epochs, see what's crackalackin
Probbbbbbably time to build a GUI for this shit and make it all nice and vaguely more useable, but... just... ugh
December 3, 2025 at 2:10 PM
Probbbbbbably time to build a GUI for this shit and make it all nice and vaguely more useable, but... just... ugh
Ok, new tokens are - if nothing else - in the tokenizer.
That... didn't take long, thank fuck for solid pipelines.
Not sure if I'm gonna start using <bos> yet - I think it'd be useful if I was putting padding on the START of the short examples to push them off into later token flow space […]
That... didn't take long, thank fuck for solid pipelines.
Not sure if I'm gonna start using <bos> yet - I think it'd be useful if I was putting padding on the START of the short examples to push them off into later token flow space […]
Original post on mastodon.social
mastodon.social
December 3, 2025 at 2:08 PM
Ok, new tokens are - if nothing else - in the tokenizer.
That... didn't take long, thank fuck for solid pipelines.
Not sure if I'm gonna start using <bos> yet - I think it'd be useful if I was putting padding on the START of the short examples to push them off into later token flow space […]
That... didn't take long, thank fuck for solid pipelines.
Not sure if I'm gonna start using <bos> yet - I think it'd be useful if I was putting padding on the START of the short examples to push them off into later token flow space […]
Alright, so - tired, and it's 1am, but this is CLEARLY the time to be adding MOAR special tokens to the dataset so the model learns how to do more markdown, like [strong]tits[/strong].
December 3, 2025 at 1:56 PM
Alright, so - tired, and it's 1am, but this is CLEARLY the time to be adding MOAR special tokens to the dataset so the model learns how to do more markdown, like [strong]tits[/strong].
December 3, 2025 at 1:37 PM
Which is a lot of words to say: I deserve another model kit.
So, feel free to enable me.
So, feel free to enable me.
December 3, 2025 at 1:25 PM
Which is a lot of words to say: I deserve another model kit.
So, feel free to enable me.
So, feel free to enable me.
Got home, ready to flop down.
But, alas, a bunch of slightly unpleasant housework needed desperate attendance (I will not gross ye with details), so now it's over an hour later and I'm more tired.
BUT, that shit is done, and done properly, and that makes me happy
But, alas, a bunch of slightly unpleasant housework needed desperate attendance (I will not gross ye with details), so now it's over an hour later and I'm more tired.
BUT, that shit is done, and done properly, and that makes me happy
December 3, 2025 at 1:24 PM
Got home, ready to flop down.
But, alas, a bunch of slightly unpleasant housework needed desperate attendance (I will not gross ye with details), so now it's over an hour later and I'm more tired.
BUT, that shit is done, and done properly, and that makes me happy
But, alas, a bunch of slightly unpleasant housework needed desperate attendance (I will not gross ye with details), so now it's over an hour later and I'm more tired.
BUT, that shit is done, and done properly, and that makes me happy
December 3, 2025 at 1:19 PM
Second #daggerheart session went well. Group's clearly flowing it better. Card juggling still diagetically broken, but... I just think that I won't do it.
December 3, 2025 at 10:58 AM
Second #daggerheart session went well. Group's clearly flowing it better. Card juggling still diagetically broken, but... I just think that I won't do it.
@MachineLordZero he's extremely sheltered and very literal, so when a girl days 'I don't have any interest in engineering' he takes that at face value rather than realising it's him being told a conversation closer.
December 3, 2025 at 3:57 AM
@MachineLordZero he's extremely sheltered and very literal, so when a girl days 'I don't have any interest in engineering' he takes that at face value rather than realising it's him being told a conversation closer.
'I understand sexism in the past was a problem'
*sighs until the sun flickers and fades*
Duuuuuuude. You have no fucking idea.
*sighs until the sun flickers and fades*
Duuuuuuude. You have no fucking idea.
December 3, 2025 at 3:56 AM
'I understand sexism in the past was a problem'
*sighs until the sun flickers and fades*
Duuuuuuude. You have no fucking idea.
*sighs until the sun flickers and fades*
Duuuuuuude. You have no fucking idea.
*long exhausted sigh* young intern telling me that engineering is male dominated because he doesn't think ladies are interested in the field because 'I talked to the ladies doing my engineering course and they don't know why other ladies don't want to do engineering'
December 3, 2025 at 3:53 AM
*long exhausted sigh* young intern telling me that engineering is male dominated because he doesn't think ladies are interested in the field because 'I talked to the ladies doing my engineering course and they don't know why other ladies don't want to do engineering'
Trusted coworker used the top sheet of my deskpad to clean the shredder and that's gonna cost her
December 3, 2025 at 3:00 AM
Trusted coworker used the top sheet of my deskpad to clean the shredder and that's gonna cost her
The Iglight is a darling of a #30MinuteMissions kit. Not the best (there's an aesthetic choice here or there I don't love) but nice, easy. A nice refreshing kit while procrastinating on going back to the RG kits
December 3, 2025 at 2:34 AM
The Iglight is a darling of a #30MinuteMissions kit. Not the best (there's an aesthetic choice here or there I don't love) but nice, easy. A nice refreshing kit while procrastinating on going back to the RG kits
Ok. #30MinuteMissions time for lunch
December 3, 2025 at 1:48 AM
Ok. #30MinuteMissions time for lunch
On one hand, last night's 'adjustments' to the LLM training code aren't on the same level as 'oh gods, I've not even been using the output of this test for nine months' like when I was dorking with LORAs, but it's still a huge shift in 'ok, everything from earlier? Nah, drop that. We're over […]
Original post on mastodon.social
mastodon.social
December 3, 2025 at 1:38 AM
On one hand, last night's 'adjustments' to the LLM training code aren't on the same level as 'oh gods, I've not even been using the output of this test for nine months' like when I was dorking with LORAs, but it's still a huge shift in 'ok, everything from earlier? Nah, drop that. We're over […]
I am two coffees and a red eye deep in today and I still don't give much of a fuck.
December 3, 2025 at 1:36 AM
I am two coffees and a red eye deep in today and I still don't give much of a fuck.
Ok, this... kinda does feel more right?
I suppose?
I suppose?
December 2, 2025 at 3:48 PM
Ok, this... kinda does feel more right?
I suppose?
I suppose?
... and it kinda feels like those times when you thought you were good, but you weren't, and now you know you weren't the root problems can be addressed?
December 2, 2025 at 2:56 PM
... and it kinda feels like those times when you thought you were good, but you weren't, and now you know you weren't the root problems can be addressed?
So, it was a hell of a shock going from these rambling messy outputs that vaguely looked like they were coalescing to:
🗣️ Prompt: Good morning, would you like coffee?
[temp: 0.8 | persona: sarcastic teen] 🤖: Good morning. Riveting stuff again later.
[temp: 1.0 | persona: sarcastic teen] 🤖: Good […]
🗣️ Prompt: Good morning, would you like coffee?
[temp: 0.8 | persona: sarcastic teen] 🤖: Good morning. Riveting stuff again later.
[temp: 1.0 | persona: sarcastic teen] 🤖: Good […]
Original post on mastodon.social
mastodon.social
December 2, 2025 at 2:55 PM
So, it was a hell of a shock going from these rambling messy outputs that vaguely looked like they were coalescing to:
🗣️ Prompt: Good morning, would you like coffee?
[temp: 0.8 | persona: sarcastic teen] 🤖: Good morning. Riveting stuff again later.
[temp: 1.0 | persona: sarcastic teen] 🤖: Good […]
🗣️ Prompt: Good morning, would you like coffee?
[temp: 0.8 | persona: sarcastic teen] 🤖: Good morning. Riveting stuff again later.
[temp: 1.0 | persona: sarcastic teen] 🤖: Good […]
Innnnnnnnnn a nutshell:
Previously, alllll my examples were padded to 1024 tokens. This meant:
* I was max training time
* This was also teaching the model to be more verbose, by default
It was also, against my wishes, shuffling the examples
* This meant, we weren't getting the 'teach long […]
Previously, alllll my examples were padded to 1024 tokens. This meant:
* I was max training time
* This was also teaching the model to be more verbose, by default
It was also, against my wishes, shuffling the examples
* This meant, we weren't getting the 'teach long […]
Original post on mastodon.social
mastodon.social
December 2, 2025 at 2:37 PM
Innnnnnnnnn a nutshell:
Previously, alllll my examples were padded to 1024 tokens. This meant:
* I was max training time
* This was also teaching the model to be more verbose, by default
It was also, against my wishes, shuffling the examples
* This meant, we weren't getting the 'teach long […]
Previously, alllll my examples were padded to 1024 tokens. This meant:
* I was max training time
* This was also teaching the model to be more verbose, by default
It was also, against my wishes, shuffling the examples
* This meant, we weren't getting the 'teach long […]
I have either broken this, really fucking badly, OR, I've fixed the lensing on it so I'm seeeing the true form of the model better?
Fucking hard to tell, innit?
Fucking hard to tell, innit?
December 2, 2025 at 2:35 PM
I have either broken this, really fucking badly, OR, I've fixed the lensing on it so I'm seeeing the true form of the model better?
Fucking hard to tell, innit?
Fucking hard to tell, innit?
On the plus side, I went chasing one thing in my pipeline and may have come out the other side having taken my training time from 16 mins down to 6?
... the fuck?
... the fuck?
December 2, 2025 at 1:46 PM
On the plus side, I went chasing one thing in my pipeline and may have come out the other side having taken my training time from 16 mins down to 6?
... the fuck?
... the fuck?