chimere
Datasets
All datasets matching “chimere”chimere-expert-predictor
Chimere Expert Predictor Models
MLP-based expert prediction models trained to prefetch MoE experts before they're needed. This is an informative negative result: despite achieving 86.65% hit@8 prediction accuracy, expert prefetch provides zero speedup on consumer hardware because CPU GEMV for small experts (430 KB) takes <50μs — less than 45% of the compute budget.
Models
Variant
Size
Description
expert_predictor/
35 MB
Base: same-layer prediction… See the full description on the dataset page: https://huggingface.co/datasets/Kevletesteur/chimere-expert-predictor.chimere-engram-tables
Chimere Engram Tables
Pre-built N-gram hash tables for the Chimere multi-tier Engram memory system. These provide O(1) lookup for domain-specific token sequences, used to bias LLM generation with factual knowledge at the logit level.
Files
File
Size
Domain
N-grams
code.engr
155 KB
Programming / software
Code patterns, API signatures
cyber.engr
172 KB
Cybersecurity
Threat indicators, CVE patterns
general.engr
679 KB
General knowledge
Common factual patterns… See the full description on the dataset page: https://huggingface.co/datasets/Kevletesteur/chimere-engram-tables.chimere-calibration
Chimere Calibration Corpus
Multi-domain calibration corpus for generating importance matrices (imatrix) for GGUF quantization. Used to produce the RAMP-v2 quantization.
Contents
5,545 samples (1.0 MB) covering: code (Python, Rust, JS, SQL), math reasoning, tool calling (BFCL-v3), French technical writing, general knowledge.
Usage
llama-imatrix -m your-model.gguf \
-f calibration_chimere.txt \
-ngl 99 --chunks 200 \
-o your-imatrix.dat
Author… See the full description on the dataset page: https://huggingface.co/datasets/Kevletesteur/chimere-calibration.chimere-quality-scores
Chimere Quality Scores
Quality evaluation data from the Chimere self-improving inference system.
Files
quality_scores.jsonl — 104 quality assessments with multi-scorer evaluation (ThinkPRM, Qwen3.5, Qwen-9B)
training_pairs.jsonl — 68 high-quality training pairs with full chain-of-thought reasoning traces
Format
Each quality score entry contains: timestamp, route, score (1-5), reasoning, verification chain-of-thought.
Each training pair contains: prompt… See the full description on the dataset page: https://huggingface.co/datasets/Kevletesteur/chimere-quality-scores.CHIME-recovery-frameworkchimere-dflash-data