Benign huggingface Download

nkkbr/Mini-K3-1H-attn-2kda-2mla-nope-v2 main

Python repeats repository clone calls

SHA-256e3d55d55ee8be21e519d570e4e6c4f55ee9a488302c1b43e5e61dd7e89b1b9de

URLhttps://huggingface.co/nkkbr/Mini-K3-1H-attn-2kda-2mla-nope-v2

MaleculeH(Po)

Evidence

English function-word token "the" lines 18–22

29… ny_for_test()
19
  # Exercise meta construction and the explicit random-materialization path.
  model = MiniK3ForCausalLM(config, device="meta")
   …

English function-word token "for" lines 22–24

31… ialize("cpu")
  kda_modules = [layer.self_attn for layer in model.model.layers if isinstance(layer.self_attn, KimiDeltaAttention)]
  mla_modules = [layer. …

Python repeats repository clone calls lines 35–53

⋯12 lines
47
  old_biases = [
      layer.feed_forward.router.e_score_correction_bias.detach().clone()
      for layer in model.model.layers
      if layer.is_moe
  ]
  model.update_quantile_balancing(output.router_rec …

English function-word token "with" lines 78–81

33… 3, 4, 5, 6, 19, 20, 21]], dtype=torch.long)
  with torch.no_grad():
      logits_a = model(input_ids=sequence_a).logits
      logits_b = model …