bonsai-2
Datasets
All datasets matching “bonsai-2”bonsai2-27b-mtp-repro
Ternary-Bonsai-2-27B + in-file MTP: reproduction bundle (RTX 4080 SUPER, Ada/SM89)
This repository holds the raw data. The method (build script, harness, launch units, write-up) lives on GitHub:
https://github.com/zhaoyilun/bonsai2-27b-mtp-repro
Both are the same piece of work: the GitHub repo has the code and the how-to, this dataset has the
measurements it produced. Cross-linked in both directions.
Raw measurements, scripts and notes for the two discussions:
official model… See the full description on the dataset page: https://huggingface.co/datasets/zhaokeqi/bonsai2-27b-mtp-repro.test-data-bonsai2bonsai2-27b-pq2-vs-fable-rtx5090
Bonsai 2 27B PQ2_0 vs Fable on a Single RTX 5090
This benchmark and reproducibility artifact compares single-GPU local LLM inference for Bonsai 2 27B / Qwen3.8 27B PQ2_0 GGUF through the Bonsai llama.cpp fork, its low-VRAM Q4_0 KV-cache quantization route with multimodal vision, and a Fable groupwise-int baseline. The controlled 210-question comparison on one RTX 5090 separates a quality-first Fable route from a 15,595 MiB sampled-peak Bonsai PQ2_0 + Q4_0-KV route.… See the full description on the dataset page: https://huggingface.co/datasets/Jerrybro/bonsai2-27b-pq2-vs-fable-rtx5090.bonsai2-dflash2-bench
Ternary Bonsai 2 27B + DFlash2 on one NVIDIA L4 — benchmark, 2026-09-24
Engine: llama-server from PrismML-Eng/llama.cpp prism @ ee8ad0ef6 + the DFlash2 cherry-pick (PrismML-Eng/llama.cpp#261),
CUDA 12.8 (sm_89). GPU: NVIDIA L4 24 GB (AWS g6.12xlarge, one server per GPU), driver 535.
Greedy (temperature 0), one request at a time, -np 1. Target GGUF prism-ml/Ternary-Bonsai-2-27B-gguf.
Drafter r3 = naklitechie/Qwen3.8-27B-DFlash2-ternary-bonsai2 Q4_K_M; stock =… See the full description on the dataset page: https://huggingface.co/datasets/naklitechie/bonsai2-dflash2-bench.Bonsai-27B-1-Bit-Distillationbonsai2-drafter-eval
Bonsai 2 drafter evaluation corpora
Prompts and greedy responses from PrismML's
Ternary-Bonsai-2-27B, recorded
as token ids together with the per-round accepted lengths of the speculative decoding loop that
produced them. The data exists to evaluate and train DFlash 2 drafters against this one target.
Target
prism-ml/Ternary-Bonsai-2-27B-mlx-2bit (MLX pack), greedy, temperature 0
Loop
mlx-dspark 0.18.0 DFlash 2 loop, stock drafter z-lab/Qwen3.8-27B-DFlash2, draft… See the full description on the dataset page: https://huggingface.co/datasets/Schiltmans/bonsai2-drafter-eval.