davidflanagancomm/minicpm2bremix2

MiniCPM 2B Remix 2

An experimental fine-tune by David Flanagan, based on the official OpenBMB MiniCPM5-2B F16 GGUF. This is a work in progress. The files are the same rank-32 AdamW LoRA update 42, merged into the base model and converted directly from the FP32 master.

Measured results

Model	S³ skill test: correct / 864	Natural finishes / 864	Math and science spot check: correct / 64
Original MiniCPM5-2B FP16	59	415	46
Original MiniCPM5-2B FP32	59	415	47
Remix 2 FP32	503	864	27
Remix 2 FP16	503	864	27
Remix 2 Q8_0	505	864	26
Remix 2 Q5_K_M	495	864	25
Remix 2 Q4_K_M	484	863	26

The S³ panel contains nine elementary arithmetic and state-tracking skills, with 96 cases per skill covering independent tasks, composition, and transfer/recovery. Scores require the requested answer format and include natural stopping behavior. The same 864 development cases were repeatedly observed during the study, so these are not a sealed final qualification or a general benchmark. The original FP32 was retested on the same runtime as the trained variants; the original FP16 score comes from the earlier runtime.

The external spot check uses the first 32 GSM8K and first 32 ARC-Challenge test rows, strict answer extraction, greedy decoding, thinking disabled, and a 512-token output limit. It is a small nonrandom sample, not an official benchmark. The FP32 remix scored 3/32 on GSM8K and 24/32 on ARC, versus 23/32 and 24/32 for the original FP32 and 22/32 and 24/32 for the original FP16. This result suggests a substantial regression on this math sample despite the S³ improvement. These scores do not establish broad quality or suitability for general use.

The original MiniCPM weights and the separately released full-model Muon remix are different from these files. Training details and research evidence: Skills³ high-rank study.