asingh15/amazon-coevolve-midtrain-pilot-sft-preview

Amazon Coevolve Mid-training SFT Preview

This is an accepted-only point-in-time snapshot of amazon-mt-pilot-native-v7 for inspecting and launching initial SFT experiments. Collection is still active, so this repository is deliberately marked incomplete.

Configurations and splits

The 50-row merged validation set holds out one complete reviewer, selected deterministically to closely match the snapshot's target-kind and construction/repair proportions. No reviewer occurs in both SFT splits. The smaller track-specific validation counts are the corresponding rows for that same reviewer.

Each row uses a portable SFT target: messages[-1].content and the top-level response are exactly <think>\n{{reasoning}}\n</think>\n{{answer}}. The component fields remain available as thinking and answer, while native_messages preserves the original Qwen-native reasoning_content representation. Only the final assistant turn is supervised via step_loss_mask=1; prompt turns have step_loss_mask=0.

from datasets import load_dataset

dataset = load_dataset("asingh15/amazon-coevolve-midtrain-pilot-sft-preview", "merged")

Snapshot status

• —Snapshot: 20260908T144909Z
• —Accepted rows: 2,611
• —Expected completed pilot rows: 8,192
• —Provisional validation: 50 rows, one reviewer, all five target kinds, both workflows
• —Candidate: Qwen/Qwen3.8-27B @ 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0
• —Teacher: deepseek-ai/DeepSeek-V4-Flash-0731 @ 7872f01b1d1fe23eabc4c98b48bffcef5a386062

See snapshot_manifest.json for split details and SHA-256 checksums, and audit/accepted_trace_index.jsonl for per-example generation metadata. Do not treat this preview validation split as the final official held-out evaluation set.