skyyyyks/demonstration_glm52

demonstration_glm52

Hindsight reasoning-chain demonstrations for SDFT, written by GLM-5.2 on the training share of the swmbench/swmbench jin10_attributed_filtered_en_daily rolling 70/30 split.

Each row is a prediction-market forecasting example; output_text is a worked reasoning chain that a skilled forecaster would have produced before knowing the outcome, ending in a residual bucket <answer>Bx</answer>. The writer was privately shown the realised move and instructed never to reveal it (a leak gate rejects hindsight phrasing).

Provenance

• —Teacher model: glm-5.2 via DashScope OpenAI-compatible endpoint, thinking disabled, temperature 0.7 (resampled at higher T on rejection).
• —Target: residual (raw delta minus frozen anchor-OLS fitted on w00_train), 9 quantile buckets B0..B8.
• —Rows: 4839 = union of all windows' train partitions (holdout backtest rows excluded).
• —Grain: hourly (3600s history/horizon), despite the source's "daily" filename.

Stats

• —with reasoning chain: 4837 / 4839 (100.0%)
• —bare-label fallback: 2
• —answer == gold bucket: 4839 / 4839 (100%)
• —rejects during generation: leak 16, format 12, transport 4

Fields

row_index, messages, label_name, label_bucket, delta, value, anchor, target_p, market_id, event_id, t, output_text

Generated with the GLM adaptation of examples/sdft/jin10-monthly/generate_demonstrations.py.