cudabenchmarktest/r7-additive-sft
Viewer • Updated • 4.04k • 44
Qwen3.5-9B R-Series Distillation
Qwen3.5-9B LoRA SFT distillation: R7 (86.8% eval) + R8 calibration. Datasets, FP16 checkpoints, and pipeline docs.
Note R7 SFT training data (4,043 samples) — Stratos + Tulu3 + SlimOrca + PrimeIntellect. Scored 86.8%.
cudabenchmarktest/r8-calibration-sft
Viewer • Updated • 7.64k • 53Note R8 SFT training data (7,635 samples) — R7 base + 7 calibration layers. Teaches 'I don't know' via R-Tuning.
cudabenchmarktest/qwen3.5-9b-r7-research
Text Generation • 9B • Updated • 474Note R7 text — FP16 safetensors. 86.8% diverse eval. Thinking + tools + instruction following.
cudabenchmarktest/qwen3.5-9b-r7-research-vision
Image-Text-to-Text • 9B • Updated • 34Note R7 vision — FP16 safetensors. Same reasoning + vision tower preserved byte-for-byte from base.
cudabenchmarktest/qwen3.5-9b-r5-research-GGUF
Text Generation • 9B • Updated • 1.2kNote R5 GGUF (superseded by R7). First round with production-quality data. 84.2% diverse eval.
cudabenchmarktest/r9-research-framework
Viewer • Updated • 1.55k • 143Note R9 research framework — policy, harness, criticality map, HERETIC integration, refusal policy + eval bucket 06. Pre-emptive release (trained R9 model pending R8b eval).