Datasets:
willtheorangeguy
/
2023-Changelog-Interviews-Transcripts

Dataset card Data Studio Files
xet
 Community
text
stringlengths
0
2.46k
[3184.12 --> 3188.08] And what it boiled down to is that there's a supposition that could you take an older model,
[3188.50 --> 3192.78] fine tune it with good data, and still actually end up getting good results?
[3192.78 --> 3198.46] With the key point being that, hey, we're only going to pay $30 to actually train the data
[3198.46 --> 3202.42] as opposed to, oh, the tens of millions of dollars that you'd have to do.
[3203.14 --> 3203.96] And could you do it?
[3203.96 --> 3204.04] Okay.
[3204.26 --> 3206.38] That was the supposition for Dolly 1.0.
[3206.62 --> 3208.50] And sure enough, we were right.
[3209.30 --> 3215.82] Basically, it was about $30 worth of training time on what is not considered public data.
[3216.04 --> 3217.20] So that's why it's Dolly 1.0.
[3217.40 --> 3217.58] Okay.
[3217.60 --> 3218.42] So we could give you the weights.
[3218.50 --> 3219.08] We could give you the model.
[3219.16 --> 3222.78] But we couldn't give you the data because the data itself was actually not public.
[3222.88 --> 3223.52] But you owned it.
[3223.94 --> 3224.24] No, no.
[3224.38 --> 3227.92] That was the, in fact, I believe it was the same data that ChatGPT was using.
[3228.12 --> 3229.30] So we could give you the weights.
[3229.30 --> 3230.30] Again, that's open source.
[3230.40 --> 3230.50] Right.
[3230.50 --> 3232.46] But we can't do the data because the data is actually ChatGPT.
[3232.54 --> 3232.74] Gotcha.
[3232.74 --> 3232.82] Okay.
[3232.82 --> 3233.42] All right.
[3233.50 --> 3238.86] So then we're going, wait, we actually used only a tiny amount of data and it still came
[3238.86 --> 3240.64] out with some pretty decent results.
[3240.84 --> 3241.04] Okay.
[3241.18 --> 3244.56] Let's go ahead and say, why don't we generate our own data?
[3245.32 --> 3247.76] So again, take credit where credit is due.
[3247.98 --> 3251.94] Our founders went ahead and said, hey, why don't we just get, we have about 5,000 employees
[3251.94 --> 3252.74] at Databricks now.
[3252.90 --> 3253.56] This is my favorite part.
[3253.68 --> 3253.82] Yeah.
[3254.00 --> 3256.22] Let's just go ahead and generate our own data.
[3256.34 --> 3258.48] So for two weeks, that's literally all we did.
[3258.48 --> 3263.88] We had basically a bunch of employees dumping in data in a Q&A style format.
[3264.02 --> 3265.12] We had seven different categories.
[3265.30 --> 3266.10] It's all listed out there.
[3266.14 --> 3267.88] So I don't remember all those details anymore.
[3268.68 --> 3269.84] I worked on the t-shirts.
[3269.98 --> 3271.30] So at least I was helpful on that part.
[3271.36 --> 3271.90] Love the t-shirt.
[3272.04 --> 3272.16] Yeah.
[3272.16 --> 3272.66] That's a good one.
[3272.74 --> 3274.42] No one's seeing this right now, but it is a...
[3274.42 --> 3274.72] Well, yeah.
[3274.72 --> 3275.44] It is a podcast.
[3275.86 --> 3276.30] That's right.
[3276.42 --> 3277.08] That tends to...
[3277.08 --> 3278.50] Draw a word picture, Adam.
[3278.50 --> 3279.30] Dude, a sheep.
[3279.52 --> 3280.34] Come on, man.
[3280.42 --> 3280.90] It's Dolly.
[3281.18 --> 3281.32] It's a sheep.
[3281.32 --> 3281.58] Dolly.
[3281.68 --> 3282.00] Dolly.
[3282.00 --> 3282.04] Dolly.
[3282.04 --> 3282.28] Sheep.
[3282.28 --> 3282.78] Oh, my gosh.
[3283.02 --> 3283.74] Oh, my goodness.
[3283.74 --> 3284.88] See, I already thought he was on point.
[3285.08 --> 3285.36] Oh.
[3286.20 --> 3286.56] Okay.
[3286.84 --> 3289.04] So Dolly, the sheep, a clone, right?
[3289.12 --> 3289.94] It's a clone, right?
[3290.02 --> 3290.88] So that's the whole context.
[3291.06 --> 3291.08] It's a clone.
[3291.34 --> 3291.58] Yes.
[3291.80 --> 3293.50] So we go ahead and actually get that up and running.
[3293.82 --> 3300.88] And then we're like, hey, now we've got 15,000 plus so set of Q&A style new information,
[3301.02 --> 3304.24] all brand new, and we're publicly giving it away, right?
[3304.24 --> 3310.42] So the actual data set, if you go to Hugging Face or Databricks Labs slash Dolly or whatever
[3310.42 --> 3314.72] the GitHub site is, basically all that data is there, okay?
[3314.84 --> 3316.02] All 15,000 lines.
[3316.46 --> 3317.56] Sorry, lines.
[3317.80 --> 3319.10] 15,000 Q&As.
[3319.34 --> 3319.58] Okay.
[3320.04 --> 3325.72] And then we train that data set again using the same old model from two years ago, okay?
[3326.10 --> 3326.38] Okay.
[3326.46 --> 3330.94] And we ran that, and then basically what was really cool about this is that it cost us
[3330.94 --> 3334.02] about $100 worth of training, but it's pretty good.
[3334.02 --> 3338.22] And if you ask some pointed questions on this stuff, it actually responds really, really
[3338.22 --> 3338.50] well.
[3338.68 --> 3343.32] For example, I've got some examples where I'm actually asking coffee questions, and the coffee
[3343.32 --> 3348.40] questions answers are, okay, I'll give ChatGBT4.0 a lot of credit.
[3348.56 --> 3348.72] Yeah.
[3348.80 --> 3352.10] It is much more verbose than what Dolly 2.0 can provide.
[3352.44 --> 3354.72] But in terms of correctness, it is correct.
[3354.84 --> 3359.34] They both are the same level of correctness between Dolly 2.0 and ChatGBT4.0.
[3359.34 --> 3363.54] I actually have it on my own, like, it's on my own GitHub somewhere, like a review where
[3363.54 --> 3364.40] I actually explain all that.
[3364.76 --> 3368.42] Mainly because I was actually running it on an M1 Mac, too, because I was goofing off and
[3368.42 --> 3369.18] decided to do it.
[3369.18 --> 3369.20] Which is fine.
[3369.34 --> 3370.30] That's amazing right there.
[3370.46 --> 3370.56] Yeah.
[3370.62 --> 3376.50] Let me first just say, as a daily user of ChatGBT, sometimes verbose is not desirable.
[3376.50 --> 3376.90] Yes.
[3376.90 --> 3381.06] And I'm like, dude, I actually will tell it to be brief or in one sentence.
[3381.16 --> 3381.64] Very specific.