llama2-13b/ is new: 40.53% reduction, KL 0.3616 -> 0.1333, MMLU 39.04% -> 49.65% against a 54.43% teacher (91.2% retained, the highest in the study). The two TT-3 control directories gain state_dict_fp16_healed.pt alongside their existing state_dict_fp16_unhealed.pt, so the ring-vs-chain comparison can be made after healing as well as at compression time. Their READMEs now carry that comparison and both confounds that qualify it. Five-task benchmarks for these checkpoints are not yet measured and every card says so.
Point every loading recipe at the renamed checkpoint
Qwen2.5-7B-Instruct compressed to 5.19B and healed: KL 0.4965 -> 0.1397, MMLU 61.67% -> 67.11% against a 75.35% teacher (89.1% retained). Apache 2.0, unlike the 3B's research licence -- the root README now states that the two Qwen sizes carry different terms. Full five-task benchmark for this checkpoint is not yet measured and the card says so.
llama2-7b/ holds the existing LLaMA-2-7B model (moved server-side, blob unchanged). qwen2.5-3b/ adds Qwen2.5-3B-Instruct compressed 32.65% and healed on 60M tokens, under the Qwen RESEARCH LICENSE (non-commercial), with LICENSE and NOTICE as that licence requires. Root README explains that licences differ per subdirectory.
