Fable-5 SFT Combined V2: 9,842 Rows for Instruction-Tuning

9,842 rows of instruction-response pairs for supervised fine-tuning, derived from two distinct Fable-5 corpora. The dataset was created by the author 'calmasacow' and was last updated on July 5, 2026. It combines 4,659 rows with added reasoning blocks and 5,183 rows of pure tool-use examples, verified to have no overlap in user-side content.

Use Cases

• Fine-tuning language models for chain-of-thought reasoning based on the 'agentic-distill-fable-5-sft' corpus with added reasoning blocks.
• Training models on tool-use interactions based on the 'fable-tool-use-sft' corpus derived from Claude Code sessions.
• Creating a combined training set for multi-task instruction-following models by merging two distinct SFT corpora.
• Studying the impact of explicit reasoning traces versus pure tool-use demonstrations on model performance.

Strengths

• Contains 9,842 total rows, a union of two distinct corpora.
• Comprises 4,659 rows with post-hoc reasoning blocks and 5,183 rows of pure tool-use examples.
• Verified to have 0% overlap between the two source corpora via SHA-256 hashing of user-side content.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known, but other metadata like file formats, size, and license are unknown.
• Data may reflect source bias inherent to the specific Fable-5 and Claude Code sessions used.

Provenance

Source

Combination of 'lordx64/agentic-distill-fable-5-sft' and 'lordx64/fable-tool-use-sft' from Hugging Face.

Collection Method

Union of two corpora, deduplicated by user-side content. Reasoning blocks were added post-hoc to one corpus by Glint-Research.

Freshness

Last updated 2026-07-05 18:19:55; freshness should be verified.