Open Issues Need Help
View All on GitHub Is the ~22% linearization ceiling a capacity wall, or a recipe limit? about 1 month ago
help wanted research question measurement
ecloudtechnology/erk-linear ecloudtechnology/erk-linear Hybrid Turkish LLM research: replacing 20% of Qwen3-14B softmax attention layers with Gated DeltaNet, with reproducible benchmarks and confidence intervals.
Open project →
0
Hybrid Turkish LLM research: replacing 20% of Qwen3-14B softmax attention layers with Gated DeltaNet, with reproducible benchmarks and confidence intervals.
Python
#efficient-attention#gated-deltanet#large-language-models#linear-attention#llm-research#open-source-ai#pytorch#qwen3#transformers#turkish-nlp
Long-context (≥16K) recall: our NIAH test saturates and cannot discriminate about 2 months ago
help wanted research question measurement
ecloudtechnology/erk-linear ecloudtechnology/erk-linear Hybrid Turkish LLM research: replacing 20% of Qwen3-14B softmax attention layers with Gated DeltaNet, with reproducible benchmarks and confidence intervals.
Open project →
0
Hybrid Turkish LLM research: replacing 20% of Qwen3-14B softmax attention layers with Gated DeltaNet, with reproducible benchmarks and confidence intervals.
Python
#efficient-attention#gated-deltanet#large-language-models#linear-attention#llm-research#open-source-ai#pytorch#qwen3#transformers#turkish-nlp
Does the placement rule generalize to other base architectures? about 2 months ago
help wanted research question measurement
ecloudtechnology/erk-linear ecloudtechnology/erk-linear Hybrid Turkish LLM research: replacing 20% of Qwen3-14B softmax attention layers with Gated DeltaNet, with reproducible benchmarks and confidence intervals.
Open project →
0
Hybrid Turkish LLM research: replacing 20% of Qwen3-14B softmax attention layers with Gated DeltaNet, with reproducible benchmarks and confidence intervals.
Python
#efficient-attention#gated-deltanet#large-language-models#linear-attention#llm-research#open-source-ai#pytorch#qwen3#transformers#turkish-nlp