CoolFace
Modelpublic

TianchenGuan/GLM-4.5-Air-USFT-unfiltered-iter1499

sourceHugging Facemitupdated 1mo agoView on Hugging Face
0likes13downloads
Model Card

GLM-4.5-Air U-SFT (unfiltered public mixture), iter 1499 (partial)

SFT of the midtrained GLM-4.5-Air on the UNFILTERED public data mixture (sft_v4_public_UNFILTERED_5050_agent2ep, 19.04M rows / 4.75M effective weight). This is the "U" control arm of a filtered-vs-unfiltered training-data-integrity experiment; the filtered twin (F-SFT, trained to completion) is GLM-4.5-Air-FSFT-public-filtered.

Training stopped at iter 1,499 of a planned 4,643 (32%) due to compute reallocation; earlier checkpoints (iter 499/999) exist. Use for matched-step comparisons against F-SFT, not as a finished model. Data recipes: see openrecipe-configs-and-ablations. Duke University research project.