aimosprite/brian-rollouts-311-specforge-turn-conversations

---
pretty_name: brian-rollouts-311-specforge-turn-conversations
license: mit
tags:
- specforge
- gpt-oss
- harmony
- conversations
- reasoning
- rollouts
- math
task_categories:
- text-generation
size_categories:
- 10K<n<100K
---

# brian-rollouts-311-specforge-turn-conversations

Turn-level `conversations` SpecForge training export derived from
`aimosprite/brian-rollouts-311-compact-turns`.

## Contents

- `brian-rollouts-311-specforge-turn-conversations.jsonl`
- `brian-rollouts-311-specforge-turn-conversations.jsonl.manifest.json`

Each JSONL row is a single assistant call rendered as a full SpecForge
`conversations` sample. The visible prefix context is intentionally repeated
across rows so that every assistant turn is independently trainable:

```json
{
"id": "amobench::amo-bench-1::attempt0::turn0",
"conversations": [
{
"role": "system",
"content": "..."
},
{
"role": "user",
"content": "..."
},
{
"role": "assistant_analysis",
"content": "..."
}
],
"dataset": "amobench",
"problem_id": "amo-bench-1",
"attempt_id": 0,
"turn_index": 0,
"turn_number": 1,
"turn_count": 18,
"rollout_date": "2026-03-11",
"source_dataset_id": "aimosprite/brian-rollouts-311-compact-turns"
}
```

## Intended usage

This export is meant for current SpecForge training flows that consume
`conversations` JSONL with the GPT-OSS chat template:

```bash
torchrun --standalone --nproc_per_node 8 scripts/train_eagle3.py \
--train-data-path ./brian-rollouts-311-specforge-turn-conversations.jsonl \
--chat-template gpt-oss
```

This turn-level export is preferable to one-row-per-rollout data for compact
GPT-OSS traces because later prompts do not retain earlier hidden `analysis`
messages. Exploding to one row per assistant call preserves the current turn's
reasoning/final output.

## Source

- Source dataset: `aimosprite/brian-rollouts-311-compact-turns`
- Conversion: token-id decoding with `openai/gpt-oss-120b`
- Rows: 69,373

---
数据集名称: brian-rollouts-311-specforge-turn-conversations
许可证: MIT许可证
标签:
- specforge
- gpt-oss
- harmony
- conversations
- reasoning
- rollouts
- math
任务类别:
- 文本生成
数据规模区间:
- 10K<n<100K
---

# brian-rollouts-311-specforge-turn-conversations

该数据集为轮次级`conversations`格式的SpecForge训练导出文件，源数据为`aimosprite/brian-rollouts-311-compact-turns`。

## 数据集内容

- `brian-rollouts-311-specforge-turn-conversations.jsonl`
- `brian-rollouts-311-specforge-turn-conversations.jsonl.manifest.json`

每个JSONL行均对应单条助手调用请求，被封装为完整的SpecForge对话样本。为确保每个助手轮次均可独立训练，所有行均会重复包含可见的前置上下文语境：

json
{
"id": "amobench::amo-bench-1::attempt0::turn0",
"conversations": [
{
"role": "system",
"content": "..."
},
{
"role": "user",
"content": "..."
},
{
"role": "assistant_analysis",
"content": "..."
}
],
"dataset": "amobench",
"problem_id": "amo-bench-1",
"attempt_id": 0,
"turn_index": 0,
"turn_number": 1,
"turn_count": 18,
"rollout_date": "2026-03-11",
"source_dataset_id": "aimosprite/brian-rollouts-311-compact-turns"
}

## 预期用途

本导出文件适配当前采用GPT-OSS对话模板处理`conversations`格式JSONL文件的SpecForge训练流程：

bash
torchrun --standalone --nproc_per_node 8 scripts/train_eagle3.py
--train-data-path ./brian-rollouts-311-specforge-turn-conversations.jsonl
--chat-template gpt-oss

相较于单轨迹单行的GPT-OSS紧凑追踪数据，该轮次级导出文件更具优势：由于后续提示无法保留此前隐藏的`analysis`字段消息，将数据拆分为单助手调用单行的形式可完整保留当前轮次的推理过程与最终输出。

## 数据源

- 源数据集：`aimosprite/brian-rollouts-311-compact-turns`
- 转换方式：通过`openai/gpt-oss-120b`进行Token-id解码
- 数据行数：69,373