Datasets:
Add TimeBraid-6.7B contextual forecasting results
Add TimeBraid-6.7B (code) to GIFT-Eval-CTX under eval/timebraid_6p7b_demo_rewrite_v1/, with History + Macro + Micro inputs. This run used the fixed 6.7B. The supplied effect JSON is rewritten as natural-language future-step conditions and convert its level/amplitude coordinates to raw units using observed-history statistics, to align with our instruction-tuning format. It is the result of a single unified model rather than an ensemble of models or an agent.
| 78-group relative MASE ↓ | Overall | Short | Medium | Long |
|---|---|---|---|---|
| TimeBraid-6.7B | 0.773281527744585 | 0.6847064982335485 | 0.8719738490515145 | 0.8640818145482951 |
The single Parquet covers all 10,000 official keys with exact finite forecast horizons and finite nonnegative MASE. Local scoring and a read-only replay through the current Space loader both reproduce the values above.
Hi team @liuxu777 , just following up on this PR. Would it be possible to review and merge it soon? We’ve verified coverage of all 10,000 official keys and confirmed that the leaderboard loader reproduces the reported scores. Please let us know if anything else is needed to move this forward. Thanks!
Hi @XinyueWangg , thanks for the contribution and for following up! We’ve reviewed the submission and will proceed with merging the PR. We’ll also update the category of TimeBraid from TSFM to TSLM (Time Series–Language Model) to reflect the method's unified time-series and language architecture. We’ll handle this metadata change on our side, so no further action is needed from you.