Datasets:
Not training ready
Hello @MoreThought , first and foremost I'd like to thank you for creating and uploading this dataset! That being said though, there is some missing or malformed data here that makes it near impossible to properly train on.
- Tool interactions are flattened into bare tool messages without tool names, arguments, IDs, or assistant tool_calls
- reasoning and final answers are not separated
- many traces contain empty assistant turns, and workflow/approval artifacts.
Please consider publishing a cleaned version with properly structured tool calls and a tool schema (row specific if different for each row), thanks again!
I appreciate your feedback, ill make sure to fix it soon.
Also, sorry for the 5 day wait.
Thanks so much! No worries on the delay, we're all busy people. Thanks again
Please keep me updated in this thread whenever you get the chance to update it, or if you don't think it will be updated. Hoping to do a new tune soon and would love to incorporate your data if it gets cleaned!
I am very sorry, I am working on a new project that requires all of my hardware, I may not update it soon.
I'll try to do both at once if possible, but if not i'll update it after.
Either way I will make sure to do it this week.
Again no pressure, just keep me posted if possible. Big thanks 🙏
same here! @armand0e does finetuning on fable 5.1 data actually improve performance?
obviously depends what you're looking for. For the most part it ends up destroying what it learned in it's later RL stages unless it's a really light SFT, but even then it adds a lot of noise to the model. It does help with style and performance in specific harnesses, it does help increase intelligence (though much less likely to when using LoRA/QLoRA compared to fft), but it does make the model drift from it's previous alignment. This is why I've been learning to make my own RL environments to help re-align the models after training them on fable data.
Overall I'd say it hurts the model if you just stop after SFT, but you can get some really good results if you take the time and compute to realign the model with GRPO and DPO afterwards.
same here! @armand0e does finetuning on fable 5.1 data actually improve performance?
obviously depends what you're looking for. For the most part it ends up destroying what it learned in it's later RL stages unless it's a really light SFT, but even then it adds a lot of noise to the model. It does help with style and performance in specific harnesses, it does help increase intelligence (though much less likely to when using LoRA/QLoRA compared to fft), but it does make the model drift from it's previous alignment. This is why I've been learning to make my own RL environments to help re-align the models after training them on fable data.
Overall I'd say it hurts the model if you just stop after SFT, but you can get some really good results if you take the time and compute to realign the model with GRPO and DPO afterwards.
It's done
Thanks! I'll take a look at it in a couple days :)
Specifically in the previously "lite" split, it is now named "train"
Not usable for training, the fn arguments all placeholders.
{
"tool_calls": [
{
...
"function": {
"name": "list_dir",
"arguments": "{\"path\": \"<directory_path>\"}"
}
Yea I haven’t done a full check yet but I noticed this too at first glance
Not usable for training, the fn arguments all placeholders.
{ "tool_calls": [ { ... "function": { "name": "list_dir", "arguments": "{\"path\": \"<directory_path>\"}" }
yeah i'm sorry but that's not really fixable without remaking the entire dataset
Not usable for training, the fn arguments all placeholders.
{ "tool_calls": [ { ... "function": { "name": "list_dir", "arguments": "{\"path\": \"<directory_path>\"}" }yeah i'm sorry but that's not really fixable without remaking the entire dataset
Then you should probably take the dataset down. You're ultimately only training models to output placeholders as a tool call. This isn't a valid dataset and it's really not ethical to make people that it is. That's not intended to be harsh because it's great that you tried this. But this will not have a good training outcome.
Not usable for training, the fn arguments all placeholders.
{ "tool_calls": [ { ... "function": { "name": "list_dir", "arguments": "{\"path\": \"<directory_path>\"}" }yeah i'm sorry but that's not really fixable without remaking the entire dataset
Then you should probably take the dataset down. You're ultimately only training models to output placeholders as a tool call. This isn't a valid dataset and it's really not ethical to make people that it is. That's not intended to be harsh because it's great that you tried this. But this will not have a good training outcome.
I understand what you are saying but the dataset is not completely useless, it is still usable for things copying the style of fable 5.1, seeing how it acts, etc...
I understand what you are saying but the dataset is not completely useless, it is still usable for things copying the style of fable 5.1, seeing how it acts, etc...
yeah, agreed. there's tons of trash out there, and of the fable 5 trace datasets I see, this is by far the highest-quality.
However, @MoreThought , I was attempting to use this set to drive an SFT for copying style and it didn't do well. I spent ages trying to troubleshoot my SFT script before I finally decided to check the dataset, and there's a bunch of weird corruption.
here are some specific examples:
Now align the research objects exactly
A good strategy would be to no ID collisions — none of the four Meraki IDs exist
To accomplish this, all six target ids are new
We should start by the four IDs are free
Writing the acceptance examine
We should start by now the showdown fixture and a generator to produce the pins
these occur hundreds of times throughout the dataset:
The approach here is to i'll start by
The key insight is that now let me
We should start by i'm
First, we need to now let me look at
The first step involves let me start by
Looking at this problem, bulk API contract pinned
The most logical first step is to next
Let's begin by now
We should start by now the
There's also a lot of weird glitched words with doubled final letters, like "baselinee", "unprotectedd", "gitignoree", and backticked words without spaces around them.
All of these seem to erode the models' sense of how sentences are formed.