fable-sft-combined-v2

# Fable-5 SFT Combined (v2) — prepared for Qwable-v2 fine-tuning 9842 rows = union of two Fable-5 SFT corpora, deduplicated by user-side content: - [`lordx64/agentic-distill-fable-5-sft`](https://huggingface.co/datasets/lordx64/agentic-distill-fable-5-sft) — 4,659 rows WITH `<think>` blocks (cleartext reasoning added post-hoc by Glint-Research) - [`lordx64/fable-tool-use-sft`](https://huggingface.co/datasets/lordx64/fable-tool-use-sft) — 5,183 rows WITHOUT `<think>` blocks (pure tool-use from 481 Fable-5 Claude Code sessions) Verified disjoint via SHA-256 on user-side content (0% overlap). The two corpora cover overlapping sessions but different rendering choices in formatting, so they contribute distinct training rows. ## Why combined for v2 Training Qwable-v2 on the union teaches the model both styles: - With-`<think>` rows reinforce reasoning-first behavior (~46% of corpus) - Without-`<think>` rows reinforce direct-action behavior (~54% of corpus) In production, the model should produce reasoning when helpful (long planning) and skip it when the action is obvious. The mixed corpus is the closest thing to that distribution we have. ## Row format Same single-`text`-column Qwen chat template as v1, but the assistant block may or may not have a `<think>...</think>` prefix depending on the source row. ## License & terms Inherits **AGPL-3.0** from both upstream corpora. Same Anthropic usage-policy considerations apply.