Datasets:

lordx64
/

fable-sft-combined-v2

license: agpl-3.0
language:
  - en
tags:
  - agentic
  - tool-use
  - distillation
  - claude
  - claude-fable-5
  - chain-of-thought
  - sft
  - qwen-chat-template
  - qwable-v2
task_categories:
  - text-generation
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.parquet

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 — 4,659 rows WITH <think> blocks (cleartext reasoning added post-hoc by Glint-Research)
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.