aimosprite/brian-rollouts-311-specforge-conversations

---
pretty_name: brian-rollouts-311-specforge-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-conversations

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

## Contents

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

Each JSONL row is one reconstructed rollout attempt in SpecForge
`conversations` format. Harmony messages are reconstructed from the compact
turn snapshots and merged so that repeated assistant text is not duplicated:

```json
{
"id": "amobench::amo-bench-1::attempt0",
"conversations": [
{
"role": "system",
"content": "..."
},
{
"role": "user",
"content": "..."
},
{
"role": "assistant_analysis",
"content": "..."
},
{
"role": "assistant_final",
"content": "..."
}
],
"dataset": "amobench",
"problem_id": "amo-bench-1",
"attempt_id": 0,
"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-conversations.jsonl \
--chat-template gpt-oss
```

This export follows SpecForge's actual Harmony loading path:
`safe_conversations_generator -> build_eagle3_dataset -> HarmonyParser`.
Consecutive `assistant_analysis` and `assistant_final` messages are preserved
as distinct Harmony messages, and assistant loss is applied by SpecForge across
assistant spans until the next `user` message.

## Source

- Source dataset: `aimosprite/brian-rollouts-311-compact-turns`
- Conversion: token-id decoding with `openai/gpt-oss-120b`
- Rows: 2,096