fable-tool-use-sft

# Fable-5 Tool-Use SFT — prepared for Qwable-v2 fine-tuning 5,183 single-turn `(user → assistant-with-tool-use)` pairs from **Claude Fable-5** (Anthropic preview model, briefly public 2026-06-10 → 2026-06-22 before being suspended globally under U.S. export-control directives), reformatted into a single-`text`-column parquet ready for `SFTTrainer(dataset_text_field="text") + train_on_responses_only`. ## Honest scope This dataset is a tool-use-focused companion to [`lordx64/agentic-distill-fable-5-sft`](https://huggingface.co/datasets/lordx64/agentic-distill-fable-5-sft) (4,659 reasoning+tool-use pairs from Glint-Research/Fable-5-traces). It's **only ~11% larger** than v1 — not a scale upgrade. The trade-off vs v1: drops the cleartext `<think>` blocks (none are usable from this source — see below) in exchange for 100% tool-use coverage and a slightly broader session distribution. ⚠️ **There may be significant overlap with v1.** Both datasets ultimately derive from Glint-Research captures of Claude Code Fable-5 sessions. SHA-256 overlap on user-side content has not been measured. Treat this as complementary to v1, not as an independent corpus. ## Why this dataset exists The 2M-row source [`Glint-Research/Complete-FABLE.5-traces-2M`](https://huggingface.co/datasets/Glint-Research/Complete-FABLE.5-traces-2M) advertises 2M rows but contains only **481 unique Claude Code sessions** — the bulk is file-snapshot / queue-op / permission-event noise inflating the row count. Pass 1 of our extractor captured all 481 sessions in the first ~50k of the 2M rows; the remaining 1.95M rows added 16 more sessions. So the source isn't 40x scale — it's the same ~hundreds of sessions with heavy event noise. Of the assistant rows actually present (~28% of the 2M): - ~28% are `thinking` blocks with **only the cryptographic signature** preserved (cleartext blocked by Anthropic's preview-model IP protection) → unusable for SFT - ~53% are real **tool_use blocks** with signature-verified Fable-5 provenance → **usable** ✓ - ~1% are cleartext `thinking` but **without signatures** → suspect provenance (likely Sonnet fallback during quota, not Fable-5) This dataset extracts the second category — real Fable-5 agentic tool calls. ## Key difference from `agentic-distill-fable-5-sft` | | `agentic-distill-fable-5-sft` (v1) | `fable-tool-use-sft` (this) | |---|---|---| | Rows | 4,659 | 5,183 | | `<think>` block in assistant turn | ✓ (cleartext, added by Glint-Research post-hoc) | ✗ (omitted — would have to be empty) | | Tool-use coverage | 81% of rows end in tool_use | 100% of rows end in tool_use | | Distribution | One developer's narrow Claude Code sessions | 481 Claude Code sessions from the FABLE.5 public window | | Unique sessions vs v1 | (reference) | likely majority-overlap — both derive from Glint captures | **Implication:** training on this dataset ALONE teaches tool-use format only; **the model will have no `<think>` reasoning capability** unless the base model already does. Use it for v2 by: - **Warm-starting** from a reasoning-capable base (e.g. [`lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled`](https://huggingface.co/lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled), same pattern as Qwable-v1) - Or **mixing** with a reasoning corpus during SFT (e.g. include `lordx64/reasoning-distill-opus-4-7-max-sft` 20-30% by weight) ## Row format ``` <|im_start|>system You are a coding agent. When you need to read, write, edit, or run code, emit XML tool calls in this exact format: <tool_use name="X" id="toolu_01abc"> {"...": "..."} </tool_use> Do NOT respond with markdown code blocks. Always use <tool_use> XML.<|im_end|> <|im_start|>user {user_or_tool_result}<|im_end|> <|im_start|>assistant {visible_text}<tool_use name="X" id="toolu_..."> {...} </tool_use><|im_end|> ``` The system prompt matches the inference recipe documented on `lordx64/Qwable-v1` — training + inference align so the model emits the correct XML format on bare prompts. ## What this repo adds on top of upstream 1. **Format**: collapsed into a single `text` column with full Qwen chat template. Drop-in for `SFTTrainer(dataset_text_field="text")`. 2. **Tool serialization**: `tool_use` blocks rendered as `<tool_use name="X" id="Y">…</tool_use>`; `tool_result` rendered as `<tool_result id="X" is_error="…">…</tool_result>`. Recoverable by a small regex (no Qwen-native `<tool_call>` tokens). 3. **Filtering**: - Dropped Claude Code's `<synthetic>` rate-limit injections - Dropped assistant turns with no tool_use (this dataset is tool-use-only) - Dropped assistant turns with empty visible content 4. **Noise stripping**: removed Claude Code slash-command meta blocks (`<local-command-caveat>`, `<command-name>`, `<command-message>`, `<command-args>`, `<local-command-stdout>`) and ANSI escape codes from user-side content. 5. **Secret scrubbing**: regex-redacted Groq/Anthropic/OpenAI/OpenRouter/HF/AWS/GitHub/Google API keys that leaked through `Read` of `.env` files in the original sessions. 6. **Dedup**: SHA-256 over user-content (drops near-duplicate retried prompts). 7. **Packaging**: single parquet under `data/train-*.parquet` for HF datasets auto-detection. Source data and capture credit go entirely to upstream — we just shaped it for training. ## Provenance chain ``` Claude Code users ──── captured raw session JSONLs during the Fable-5 (multiple authors, public window (2026-06-10 → 2026-06-22), before via `1EYE4ALL` capture) Anthropic suspended Fable-5 globally │ ▼ 1EYE4ALL/Fable-5-traces ──── HF dataset of raw events │ ▼ Crownelius/Complete- ──── deduplicated + cleaned (session-limit FABLE.5-traces-2M answer-pattern rows removed) │ ▼ Glint-Research/Complete- ──── re-host with provenance metadata FABLE.5-traces-2M (`row_hash`, `first_source_dataset`, etc.) │ ▼ lordx64/fable-tool-use-sft ──── this dataset: extract (user, asst-with-tool-use) pairs, Qwen-template, scrub, dedup, push ``` Authenticity is verified via the cryptographic signatures on the (redacted) thinking blocks — those signatures are valid Anthropic API output, unforgeable, and contain the `claude-fable-5` model_id when decoded. The tool_use blocks share the same signed envelope. ## License & terms Inherits **AGPL-3.0** from the upstream `Glint-Research/Complete-FABLE.5-traces-2M` redistribution. Downstream users running models trained on this dataset in a network-accessible service must comply with AGPL §13 (source disclosure for network use). The underlying content is output from Anthropic's gated `claude-fable-5` preview model. Downstream users should verify compliance with [Anthropic's usage policies](https://www.anthropic.com/legal/usage-policy) for their specific use case before fine-tuning further or building commercial products.