Fable v2 is out, and the tool-call split is worth a swap

#1
by saidutta69 - opened

Congrats on Forge-1 - 6k downloads and 27 likes is better than most fine-tunes built on this dataset get.

One thing worth a re-pull: this card points at fable-5-premium, and there are two newer cuts now.

  • fable-5-premium-v2 - same traces, cleaned.
  • fable-5.1-premium-v2 - adds a claude_code split with real tool-call values: actual file paths and shell commands, rather than whole-value placeholders like <file_path> / <bash_command>.

If Forge-1 was trained on the code split rather than claude_code, that's the difference between a model that knows when to call Edit and one that knows what to put in it. Cheap experiment on a 2B: swap the split, hold everything else fixed, and watch tool-name and argument validity.

Happy to tell you which split I'd pick for a coding-agent mix if you share the current recipe, and to compare notes on eval numbers - I have a LFM3 tool-calling fine-tune on the same data, so the comparison would be apples to apples.

Either way, nice work on it.

YoSaNoo pinned discussion

Sign up or log in to comment