Caphlon-Model
A monolingual English large model obtained through a three-step compression process of vocabulary slimming, layer pruning, and low-bit dynamic quantization. It is output as a single GGUF file and is intended for local deployment on consumer-grade hardware.
Timeline (last 90 days)
The construction of Caphlon-Model is divided into three steps: vocabulary surgery, layer pruning, and quantization.
Vocabulary surgery reduces the merged vocabulary to a target of 100k, saving about 649 MiB, at the cost of making the model strictly English monolingual.
Layer pruning removes 4 redundant layers, layers 12 to 15, saving about 302 MiB.
Quantization uses the Unsloth Dynamic IQ2_S format, resulting in a final single-file size of 6.87 GiB.
Evaluated with llama.cpp on a Tesla T4, the HumanEval pass@1 retention rate is 96.7% (29/30).
Evaluated with llama.cpp on a Tesla T4, the MMLU subset retention rate is 92.9% (344/502).
English perplexity is 6.6088, compared with the baseline of 6.3638.
These benchmark results are currently self-reported from a single source, the evaluation sets are limited in size (30 HumanEval questions, a 502-question MMLU subset), and they have not yet been independently reproduced and cross-validated by the community.
Caphlon-Model is positioned as English-only, low-memory, and single-file, suitable for resource-constrained local inference and edge device experiments.
The author has released the model and code on GitHub, the project website, and Hugging Face (demiralpdev/Caphlon-Model, caphlon.dev, ssuasd/Caphlon-Model).
All Facts (10)
The construction of Caphlon-Model is divided into three steps: vocabulary surgery, layer pruning, and quantization.
50%UnverifiedVocabulary surgery reduces the merged vocabulary to a target of 100k, saving about 649 MiB, at the cost of making the model strictly English monolingual.
50%UnverifiedLayer pruning removes 4 redundant layers, layers 12 to 15, saving about 302 MiB.
50%UnverifiedQuantization uses the Unsloth Dynamic IQ2_S format, resulting in a final single-file size of 6.87 GiB.
50%UnverifiedEvaluated with llama.cpp on a Tesla T4, the HumanEval pass@1 retention rate is 96.7% (29/30).
50%UnverifiedEvaluated with llama.cpp on a Tesla T4, the MMLU subset retention rate is 92.9% (344/502).
50%UnverifiedEnglish perplexity is 6.6088, compared with the baseline of 6.3638.
50%UnverifiedThese benchmark results are currently self-reported from a single source, the evaluation sets are limited in size (30 HumanEval questions, a 502-question MMLU subset), and they have not yet been independently reproduced and cross-validated by the community.
50%UnverifiedCaphlon-Model is positioned as English-only, low-memory, and single-file, suitable for resource-constrained local inference and edge device experiments.
50%UnverifiedThe author has released the model and code on GitHub, the project website, and Hugging Face (demiralpdev/Caphlon-Model, caphlon.dev, ssuasd/Caphlon-Model).
50%