MiniCPM5-2B-EN-Strict-i1-GGUF (private)

imatrix GGUF quants of adityabhushannagar/MiniCPM5-2B-EN-Strict (vocabulary-trimmed MiniCPM5; see that repo for the trimming details, removal manifest and verification).

• imatrix: mradermacher's MiniCPM5-2B.imatrix.gguf (from mradermacher/MiniCPM5-2B-i1-GGUF, computed on the original model). It is reused unchanged: trimming only removes vocab rows, so every linear layer and its input activations are identical, and output.weight is keyed by its (unchanged) input width.
• Source for quantization: BF16 GGUF converted from the bf16 HF weights (lossless source). Quant types: the mradermacher i1 naming (i1-<TYPE>); only a subset of mradermacher's 24 types was built.
• Tool: llama.cpp build b11457 (prebuilt CPU binaries). convert_hf_to_gguf.py needed a 3-line patch to map the trimmed tokenizer fingerprint to the native minicpm5 pre-tokenizer (same regex as the original), otherwise it refuses unknown tokenizers.
• Variant: STRICT (ASCII-only vocab, no non-ASCII byte tokens). In llama.cpp there is no HF normalizer, so pre-fold input to ASCII yourself (e.g. unidecode); bytes >=0x80 have no token and are silently dropped.

Tests (CPU, 1 thread, same text for all rows)

| file | size GB | PPL (2x512 tok, ASCII wikitext-2) | non-ASCII tokenize rc | greedy sample ("Q: What is 17 + 25? ...") |

|---|---|---|---|---|

| MiniCPM5-2B-EN-Strict.i1-Q6_K.gguf | 1.93 | 12.7637 +/- 1.81700 | 0 | 42. Q: What is 17 + 25? |

| MiniCPM5-2B-EN-Strict.i1-Q5_K_M.gguf | 1.68 | 12.9058 +/- 1.85642 | 0 | |

| MiniCPM5-2B-EN-Strict.i1-Q4_K_M.gguf | 1.44 | 13.0759 +/- 1.86571 | 0 | 42. Q: What is the capital of France? Answer |

| MiniCPM5-2B-EN-Strict.i1-IQ3_M.gguf | 1.12 | 14.9146 +/- 2.11569 | 0 | 42. Q: What is 17 + 25? |

| MiniCPM5-2B-EN-Strict.i1-IQ2_M.gguf | 0.88 | 34.3981 +/- 5.55905 | 0 | 25 + 17 = 42. If you want |

| MiniCPM5-2B-EN-Strict.bf16.gguf | nan | 12.8382 +/- 1.83704 | 0 | 42. Q: What is 17 + 25? |

Reference, same protocol on mradermacher's ORIGINAL (untrimmed) i1 GGUFs:

ref2/MiniCPM5-2B.i1-Q4_K_M.gguf | PPL=13.0378 +/- 1.85730 
ref2/MiniCPM5-2B.i1-IQ2_M.gguf | PPL=35.4807 +/- 5.72363

PPL is on a small text (1,024 tokens) so absolute values are noisy; compare rows with each other, not with other benchmarks.

llama.cpp tokenization of the text café 你好 😀 ok (no HF normalizer in llama.cpp; bytes without a token are dropped):

     0 -> '<s>'
    88 -> 'c'
  1934 -> 'af'
   148 -> ' '
   148 -> ' '
 18081 -> ' ok'