TraceForge-Fable-5-Reasoning-Cursor

Technical report: TraceForge: Distilling with Your Own Thinking

The original Fable 5 coding-agent trajectories are derived from TeichAI/Fable-5-Cursor-Traces, published by TeichAI with data contributed by Enderchef. This release contains hundreds of Cursor + Fable 5 coding-agent trajectories, presented as assistant-turn views with their preceding history.

Yan Wang · Independent Researcher

Technical report (PDF) · 中文说明

TraceForge takes a strong teacher's answer and generates an accompanying reasoning trace in the style of the student's own model family. The generated trace and the teacher's answer can then be paired as SFT data. This release lets readers inspect the resulting data, alongside the original reasoning-bearing trajectories.

Technical report

Figure 1 — Learning from different reasoning traces

Figure 1 compares coding performance on HumanEval+ and MBPP+ as training proceeds. With Fable 5's answers held fixed, replacing its original reasoning with TraceForge's family-native reasoning improves the final scores from 64.63 to 68.29 on HumanEval+ and from 52.38 to 59.52 on MBPP+. The MBPP+ lead holds at every displayed checkpoint. The other curves show TraceForge paired with SOL-6.1 answers, Qwen3.8-max's own data, and TraceInversion paired with SOL-6.1 answers. They illustrate how changing the source of the reasoning trace and answer changes the student's learning trajectory.

Figure 2 — How family-native reasoning is generated

The generator receives the conversation context, including user messages, tool calls and responses, and the known answer, with the reasoning trace removed. It learns to fill in the missing reasoning from genuine traces written by the student's model family. A critic learns to distinguish genuine family traces from generated ones and supplies adversarial feedback to improve the generator; it can also select among generated candidates. The resulting family-native reasoning gives a derivation for the teacher's answer in a form familiar to the student's family.

Original trajectories: Cursor + Fable 5

original-trajectories/data.jsonl.gz contains Cursor + Fable 5 trajectories with their original reasoning, answers/actions, conversation context and available tool information. Each row is an assistant-turn view with its preceding history.

Each assistant message now includes a string field, Family-native content:

• content retains that message's existing content.
• Family-native content contains the complete alternative content: the generated family-native reasoning inside <think>...</think>, followed by the unchanged original answer/action. Only the reasoning within the think block is replaced; the text after </think> is preserved exactly.
• It is "" when no corresponding generated reasoning is available. It never falls back to teacher reasoning.

Existing generated reasoning is matched to assistant messages by source trajectory and message index. When the same assistant turn appears in another row's history, its available alternative content is attached to that message too. User and tool messages retain their original structure. Answers/actions and structured tool_calls are retained.

The field is on the assistant message it belongs to, so there is no top-level reasoning field to match to a turn. The example below is schematic; the historical assistant's existing content can be empty while its action is represented in tool_calls.

{
  "id": "example-assistant003",
  "messages": [
    {
      "role": "user",
      "content": "修复 calc.py 中的加法函数。"
    },
    {
      "role": "assistant",
      "content": "",
      "tool_calls": [
        {
          "id": "read_1",
          "type": "function",
          "function": {
            "name": "read_file",
            "arguments": "{\"path\":\"calc.py\"}"
          }
        }
      ],
      "Family-native content": "<think>先查看当前实现，确认错误位置。</think>\n读取 calc.py。"
    },
    {
      "role": "tool",
      "tool_call_id": "read_1",
      "content": "def add(a, b):\n    return a - b"
    },
    {
      "role": "assistant",
      "content": "<think>函数名表示加法，但实现用了减法，需要修正运算符。</think>\n将 return a - b 修改为 return a + b。",
      "Family-native content": "<think>add 应返回两数之和，当前代码返回两数之差，因此把减号换成加号。</think>\n将 return a - b 修改为 return a + b。"
    }
  ]
}

In the final assistant message above, both fields end with the same edit instruction; only the reasoning differs. To read the alternative target for the last turn, use row["messages"][-1]["Family-native content"].

If no generated reasoning is available, the new field is an empty string:

{
  "role": "assistant",
  "content": "<think>原思考...</think>\n原答案或动作...",
  "Family-native content": ""
}

A non-empty original content does not imply that generated reasoning is available. Among the 7,094 target assistant turns, 4,854 have generated reasoning and 2,240 do not; the latter have an empty Family-native content.

Source: TeichAI/Fable-5-Cursor-Traces. The original reasoning and answers/actions come from the upstream trajectories; Family-native content is the TraceForge-generated alternative added in this release.

Five SFT datasets

The five collections below are already in SFT format, each consolidated into one folder. Files are gzip-compressed UTF-8 JSONL, with one sample per line.

Collection	Contents	File
Fable 5 data	Original Fable 5 reasoning and answers/actions.	data.jsonl.gz
TraceForge reasoning + Fable 5 answer	TraceForge family-native reasoning paired with Fable 5 answers/actions.	data.jsonl.gz
Qwen3.8-max data	Qwen3.8-max reasoning and answers.	data.jsonl.gz
TraceInversion + SOL-6.1 answer	TraceInversion-generated reasoning paired with SOL-6.1 answers/actions.	data.jsonl.gz
TraceForge reasoning + SOL-6.1 answer	TraceForge family-native reasoning paired with SOL-6.1 answers/actions.	data.jsonl.gz

For each row, use messages[:-1] as the context and messages[-1].content as the supervised target:

messages = row["messages"]
context = messages[:-1]
target = messages[-1]["content"]