---
license: agpl-3.0
pretty_name: Fable 5 Traces (messages format)
task_categories:
- text-generation
tags:
- agent-traces
- claude-code
- fable-5
- chain-of-thought
- tool-use
- coding-agents
- synthetic-data
- distillation
- sharegpt
- unsloth
---

# Fable 5 Traces — `messages` format

A conversion of [Glint-Research/Fable-5-traces](https://huggingface.co/datasets/Glint-Research/Fable-5-traces)
(`fable5_cot_merged.jsonl`) into the standard HF/OpenAI **conversational `messages` format**, ready to drop
into Unsloth or TRL `SFTTrainer` notebooks with no extra conversion step.

## Format

Each row:

```json
{
  "messages": [
    {"role": "user", "content": "<flattened prior transcript / context>"},
    {"role": "assistant", "content": "<think>...cot...</think>\nASSISTANT (tool call) Read input={...}"}
  ],
  "uid": "source session id#index",
  "session": "source session id",
  "source_file": "original raw trace path",
  "model": "claude-fable-5",
  "output_type": "tool_use | text",
  "origin": "local | hf"
}
```

The assistant turn is the original `completion` field from the source dataset: a `<think>...</think>`
reasoning block (from `cot`) followed by either a tool call (`ASSISTANT (tool call) <Tool> input={...}`)
or a plain assistant message (`ASSISTANT (message): ...`).

## Loading with Unsloth / TRL

```python
from datasets import load_dataset

dataset = load_dataset("rex099/fable-5-traces-messages", split="train")
# dataset[0]["messages"] -> [{"role": "user", ...}, {"role": "assistant", ...}]

# TRL's SFTTrainer auto-detects the "messages" column and applies the
# tokenizer's chat template automatically:
from trl import SFTTrainer, SFTConfig
trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    train_dataset=dataset,
    args=SFTConfig(dataset_text_field=None, ...),
)
```

## Provenance

Derived entirely from the merged training rows of `Glint-Research/Fable-5-traces`
(`fable5_cot_merged.jsonl`), reshaped 1:1 into `messages` — no content was altered.
See the source dataset card for full details on collection, fields, and caveats
(license: AGPL-3.0, coding-agent telemetry — not sanitized benchmark data).