A collection of LogitsProcessors to customize and enhance LLM behavior for specific tasks.
-
Updated
Sep 16, 2026 - Python
#
logits
Here are 28 public repositories matching this topic...
Research design for a Jev-native agent system:enable more options than jev provided with virtulization and paging, tool integration, external helper logits Top-k proposals with Jev-controlled fallback ,decision-aware hierarchical memory, and dependency-aware replanning.
research system virtualization top-k typesafe natural-language-generation ai-agents paged logits technical-design function-calling small-language-models tool-calling agent-memory diffusion-language-models decision-models hierarchical-memory jev paged-options
-
Updated
Sep 28, 2026 - Python
Typed decisions (choice / score / yes-no) from local Qwen models on Apple Silicon. Probabilities come straight from the logits, no text generation. TypeSafe-compatible HTTP API, runs on MLX.
decision-making classification mlx structured-output fastapi logits apple-silicon zero-shot-classification llm local-llm qwen llm-evaluation
-
Updated
Sep 21, 2026 - Python
Plots how the logit values that are passed into the softmax function change over time as the model is trained.
-
Updated
Aug 23, 2020 - Python
Jev-style typed decisions from Qwen3.5-2B logits — one forward pass, zero decoding, zero fine-tuning.
transformers self-hosted fastapi logits zero-shot-classification qwen qwen3-5 jev system-one typed-decisions
-
Updated
Sep 22, 2026 - Python
Novel Hallucination detection method
-
Updated
Jul 2, 2026 - Python
Local frozen-backbone decision readout: evidence + criterion + options in, a probability per option out, in one forward pass. JevBench public set 0.805 / hard 0.604 with Qwen3.5-4B, no training.
-
Updated
Sep 24, 2026 - Python
Tell recoverable LLM failures from structural ones by reading the failed trace. Two models, one vLLM pass, no extra forward pass.
reasoning interpretability fine-tuning logits vllm llm-inference llm-evaluation test-time-compute logit-processor proxy-tuning
-
Updated
Aug 19, 2026 - Python
Convert any causal LM into a Jev-style typed decision model — no training, no new weights. Measured honestly against JevBench, negative results included.
-
Updated
Sep 24, 2026 - Python
Detecting prompt toxicity from a small LLM's internal logits and embeddings, no separate moderation model needed. 99.18% AUPRC with a 64-neuron MLP.
nlp machine-learning text-classification transformers pytorch embeddings mlp toxicity-detection logits llm
-
Updated
Jul 11, 2026 - Jupyter Notebook
Typed decisions from a local LLM in one forward pass. Reads the logits of the allowed options instead of generating text. ~44 ms per extra decision on the same document with llama.cpp prefix caching.
inference classification zero-shot fastapi latenc logits llm llama-cpp qwen prefix-caching system-one
-
Updated
Sep 26, 2026 - HTML
Temporal analysis of LLM safety activation via logit-margin scores.
python alignment jailbreak-detection explainable-ai logits llm-safety ai-safety-research mechanistic-analysis
-
Updated
Jul 20, 2026 - Python
Fast, System One-inspired local AI decisions without text generation. One E4B Tetris snapshot game averaged 33.3 ms per request on an RTX 5090. Open source.
-
Updated
Sep 28, 2026 - Python
Ask a vision model a question, get the answer from its logits in one forward pass. No decoding, nothing trained. The confidence is the logit gap in nats.
computer-vision logits zero-shot-classification open-vocabulary vision-language-model llm-inference qwen
-
Updated
Sep 22, 2026 - Python
Composable token sampling over f32 logits with caller-supplied randomness
-
Updated
Aug 31, 2026 - Rust
Multi strategy logit fusion between Ouro-1.4B (Universal Transformer) and HRM-Text-1B (prefix-LM) for improved text generation
entropy transformers fusion benchmarks language-model kl-divergence perplexity universal-transformer logits llm ouro-1-4b hrm-text-1b js-div
-
Updated
Jul 1, 2026 - Python
Convolutional Networks Training and Logits Saving
-
Updated
May 6, 2020 - Python
This project bridges the gap between human language and machine-executable code by translating prompts into structured function calls with typed arguments. Built strictly with Python 3.10+ and Pydantic , it demonstrates how constrained decoding can ensure near-perfect reliability and strict JSON compliance, even on small 0.5B parameter models.
-
Updated
Jul 13, 2026 - Python
Add this topic to your repo
To associate your repository with the logits topic, visit your repo's landing page and select "manage topics."