aimosprite/brian-rollouts-311-specforge-conversations

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:

{
  "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:

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