DOCUMENTATION • SPECIFICATIONS • TOOLING

Documentation Hub

Complete guides for inspecting Solace distillation datasets, deploying sub-8B quantized weights locally, and running fine-tuning harnesses.

$anvil run huihui-qwen --ctx 262144 --type-k turbo4 --type-v turbo3
DATASETS & TRACES2 guides

Multi-Teacher Reasoning Corpora

Schema specifications, Apache Parquet downloads, and Hugging Face dataset loaders.

Project Solace 1.0 Omni Dataset GuideHow to inspect, download, filter, and train models on the 12.59M verified frontier conversation corpus (137.7 GB).
→
Dataset Schema & Parquet SpecificationDetailed Apache Parquet field definitions, reasoning token tags, and multi-teacher provenance metadata.
→
STUDENT WEIGHTS1 guide

Sub-8B Quantized Checkpoints

vLLM, llama.cpp, Ollama, and Apple MLX deployment guides for FP8, INT4 AWQ, and GGUF.

Quantized Models & Anvil QuickstartStep-by-step recipes for pulling and running Solstice quantized checkpoints with the Anvil CLI and Google TurboQuant KV cache compression.
→
VERIFICATION HARNESS2 guides

Rejection Sampling & Tooling

Automated compiler sandboxes, Lean 4 theorem verifiers, and consensus ranking pipelines.

Multi-Teacher Verification & Rejection Sampling PipelineHow Solstice-AI filters, validates, and aligns reasoning traces using compiler sandboxes and mathematical assertion engines.
→
Anvil: TurboQuant KV Cache & Speculative Decoding EngineDeploying Google TurboQuant FWHT-rotated KV cache compression, single-word model launches, and high-throughput inference with the Anvil CLI.
→
HOW-TO RECIPES1 guide

Fine-Tuning & Training Recipes

End-to-end recipes for fine-tuning Llama 3.1 and Qwen 2.5 on consumer and cloud GPUs.

Fine-Tuning Student Models with Solace 1.0Step-by-step recipe for fine-tuning sub-8B base models (Llama 3.1 8B, Qwen 2.5 7B) using Axolotl, Unsloth, and TRL.
→