Text Generation • 36B • Updated • 319 • 104
text stringlengths 673 80.8k | source stringclasses 1
value | session stringclasses 1
value |
|---|---|---|
<|im_start|>system
You are a helpful AI assistant.<|im_end|>
<|im_start|>user
USER: Make a new one, it should be a fast paced multiplayer FPS (with optinal bots) with full ray-traced stuff with lots of settings. TL;DR clone CSgo to the best of your ability. Look at the result and refine it
USER: USER: <command-name>/mo... | Glint-Research/Fable-5-traces | glint |
<|im_start|>system
You are a helpful AI assistant.<|im_end|>
<|im_start|>user
USER: Make a new one, it should be a fast paced multiplayer FPS (with optinal bots) with full ray-traced stuff with lots of settings. TL;DR clone CSgo to the best of your ability. Look at the result and refine it
USER: USER: <command-name>/mo... | Glint-Research/Fable-5-traces | glint |
<|im_start|>system
You are a helpful AI assistant.<|im_end|>
<|im_start|>user
USER: Make a new one, it should be a fast paced multiplayer FPS (with optinal bots) with full ray-traced stuff with lots of settings. TL;DR clone CSgo to the best of your ability. Look at the result and refine it
USER: USER: <command-name>/mo... | Glint-Research/Fable-5-traces | glint |
<|im_start|>system
You are a helpful AI assistant.<|im_end|>
<|im_start|>user
USER: Make a new one, it should be a fast paced multiplayer FPS (with optinal bots) with full ray-traced stuff with lots of settings. TL;DR clone CSgo to the best of your ability. Look at the result and refine it
USER: USER: <command-name>/mo... | Glint-Research/Fable-5-traces | glint |
<|im_start|>system
You are a helpful AI assistant.<|im_end|>
<|im_start|>user
USER: Make a new one, it should be a fast paced multiplayer FPS (with optinal bots) with full ray-traced stuff with lots of settings. TL;DR clone CSgo to the best of your ability. Look at the result and refine it
USER: USER: <command-name>/mo... | Glint-Research/Fable-5-traces | glint |
<|im_start|>system
You are a helpful AI assistant.<|im_end|>
<|im_start|>user
USER: Make a new one, it should be a fast paced multiplayer FPS (with optinal bots) with full ray-traced stuff with lots of settings. TL;DR clone CSgo to the best of your ability. Look at the result and refine it
USER: USER: <command-name>/mo... | Glint-Research/Fable-5-traces | glint |
<|im_start|>system
You are a helpful AI assistant.<|im_end|>
<|im_start|>user
USER: Make a new one, it should be a fast paced multiplayer FPS (with optinal bots) with full ray-traced stuff with lots of settings. TL;DR clone CSgo to the best of your ability. Look at the result and refine it
USER: USER: <command-name>/mo... | Glint-Research/Fable-5-traces | glint |
<|im_start|>system
You are a helpful AI assistant.<|im_end|>
<|im_start|>user
USER: Make a new one, it should be a fast paced multiplayer FPS (with optinal bots) with full ray-traced stuff with lots of settings. TL;DR clone CSgo to the best of your ability. Look at the result and refine it
USER: USER: <command-name>/mo... | Glint-Research/Fable-5-traces | glint |
<|im_start|>system
You are a helpful AI assistant.<|im_end|>
<|im_start|>user
…[earlier truncated]…
-x--- 161 lane lane 20480 Jun 12 14:08 ..
drwxrwxr-x 3 lane lane 4096 Jun 12 09:49 archives
drwxrwxr-x 86 lane lane 4096 Jun 12 09:52 node_modules
-rw-rw-r-- 1 lane lane 287 Jun 12 09:52 package.json
-rw-rw-r-- ... | Glint-Research/Fable-5-traces | glint |
<|im_start|>system
You are a helpful AI assistant.<|im_end|>
<|im_start|>user
…[earlier truncated]…
atform — auth, avatars, multiplayer worlds, studio editor, real-time physics — generated in a single shot. Express backend, Three.js 3D rendering, WebSocket multiplayer. 4 pre-built official games, catalog system, friend... | Glint-Research/Fable-5-traces | glint |
<|im_start|>system
You are a helpful AI assistant.<|im_end|>
<|im_start|>user
…[earlier truncated]…
1 const express = require('express');
2 const path = require('path');
3 const fs = require('fs');
4 const { spawn } = require('child_process');
5
6 const app = express();
7 const PORT = Number(process.env.PORT) || 4000... | Glint-Research/Fable-5-traces | glint |
<|im_start|>system
You are a helpful AI assistant.<|im_end|>
<|im_start|>user
…[earlier truncated]…
l ray-traced stuff with lots of settings. TL;DR clone CSgo to the best of your ability. Look at the result and refine it
USER: USER: <command-name>/model</command-name>
<command-message>model</command-message... | Glint-Research/Fable-5-traces | glint |
<|im_start|>system
You are a helpful AI assistant.<|im_end|>
<|im_start|>user
…[earlier truncated]…
d-caveat>
USER: <command-name>/model</command-name>
<command-message>model</command-message>
<command-args></command-args>
USER: <local-command-stdout>Set model to Fable 5 and saved as your defaul... | Glint-Research/Fable-5-traces | glint |
<|im_start|>system
You are a helpful AI assistant.<|im_end|>
<|im_start|>user
…[earlier truncated]…
lane lane 4096 Jun 12 09:52 node_modules
-rw-rw-r-- 1 lane lane 287 Jun 12 09:52 package.json
-rw-rw-r-- 1 lane lane 37037 Jun 12 09:52 package-lock.json
drwxrwxr-x 4 lane lane 4096 Jun 12 09:49 public
drwxrwx... | Glint-Research/Fable-5-traces | glint |
<|im_start|>system
You are a helpful AI assistant.<|im_end|>
<|im_start|>user
…[earlier truncated]…
-type f | grep -v node_modules | head -40", "description": "Inspect package.json and file tree"}
TOOL RESULT: {
"name": "ai-archives",
"version": "1.0.0",
"description": "AI Archives — only the leaps that matter",... | Glint-Research/Fable-5-traces | glint |
End of preview. Expand in Data Studio
Fable-5 SFT — prepared for Qwable fine-tuning
4,659 single-turn pairs from Claude Fable-5 (Anthropic preview model, suspended globally 2026-06-22 under U.S. export-control directives), reformatted into a single-text-column parquet ready for SFTTrainer(dataset_text_field="text") + train_on_responses_only.
Composition:
- 3,793 rows (81%) end in a
<tool_use>block — agentic tool-call patterns - 866 rows (19%) end in a pure text response
This is agentic data, not pure reasoning data. The <think> blocks reason about code edits and tool calls, not about abstract problems. The narrow training distribution is essentially one developer''s week of Claude Code sessions plus assorted preview-tool work.
Used to train
🤖 lordx64/Qwable-v1 — Qwen3.6-35B-A3B warm-started from the Opus 4.7 distill, then SFT''d on this dataset for 2 epochs (582 steps) on a single H200. Final loss 0.7956 (last-20 avg).
GGUF quants: lordx64/Qwable-v1-GGUF (IQ4_XS, Q5_K_M, Q8_0).
What this repo adds on top of upstream
- Format: collapsed into a single
textcolumn with full Qwen chat template (system + user +<think>…</think>+ visible response). - Tool serialization:
tool_useblocks rendered as<tool_use name="X" id="Y">…</tool_use>;tool_resultrendered as<tool_result id="X">…</tool_result>. Custom XML envelope — NOT Qwen native<tool_call>tokens. Parses with a small regex. - Filtering: dropped Claude Code''s
<synthetic>rate-limit injections; dropped assistant turns with no thinking trace; dropped rows where CoT < 50 chars or output failed validation. - 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. - Secrets scrubbed: 204 occurrences of 2 active Groq API keys redacted from upstream session JSONLs (Claude Code captured them via
Readof.envfiles in the original developer''s repo). - Dedup: SHA-256 over user-turn content (6 dups removed).
- Packaging: single parquet under
data/train-*.parquetfor HF datasets auto-detection.
The source data and CoT credit go entirely to the upstream creators — we just shaped it for training.
Provenance
| Source | Rows | Chars |
|---|---|---|
Glint-Research/Fable-5-traces |
4,659 | 48,772,554 |
Total: 48,772,554 chars (~12,193,138 Qwen tokens), avg ~2.6k tokens/row.
Note on the other Fable-5 sources
We initially planned a 3-source merge including armand0e/claude-fable-5-claude-code and victor/fable-5-boeing-747-trace. Both turned out to have 100% of thinking blocks redacted by Anthropic''s API (only the cryptographic signature is preserved; the cleartext thinking field is empty). This is consistent with Anthropic''s preview-model IP protection — useful for trace integrity verification, but not for SFT.
Only Glint-Research/Fable-5-traces ships cleartext CoT (their README notes they "added the CoT data" — likely they had a different API tier or synthesized post-hoc; quality is high-signal reasoning consistent with the final response).
Row format
<|im_start|>system
You are a helpful AI assistant.<|im_end|>
<|im_start|>user
{user_or_tool_result}<|im_end|>
<|im_start|>assistant
<think>
{fable5_thinking}
</think>
{fable5_response_with_tool_calls}<|im_end|>
License & terms
Inherits AGPL-3.0 from upstream Glint-Research/Fable-5-traces. The underlying content is output from Anthropic''s gated claude-fable-5 model (briefly available 2026-06-10 to 2026-06-22) — downstream users must verify compliance with Anthropic''s usage policies for their specific use case.
- Downloads last month
- 128