kld
Datasets
All datasets matching “kld”kimi-k3-full-mxfp4-kld-reference-32x2048
Kimi K3 full-MXFP4 KLD reference logits
This dataset contains the canonical full-vocabulary reference logits for
quantization comparisons of Kimi K3. The source is the original full MXFP4
checkpoint served as W4A16 on TP16 with vLLM dev/gg-k3, SparkInfer, and
InstantTensor.
Contents
32 independent 2048-token windows
65,504 scored next-token positions (32 * 2047)
vocabulary size 163,840
one [2047, 163840] F32 safetensors tensor per window
tensor key: logits
total… See the full description on the dataset page: https://huggingface.co/datasets/festr2/kimi-k3-full-mxfp4-kld-reference-32x2048.GLM-5.2-Unsloth-Style-KLD-Refs-20260707
GLM-5.2 Unsloth-Style KLD BF16 Reference Logits
This dataset contains the BF16 reference prompt logprobs used for the
GLM-5.2 Unsloth-style prefill KLD run measured on 2026-07-07.
The reference cache is intended to make the candidate-side KLD comparison
reproducible without rerunning the expensive BF16 reference pass.
Reference Logits
reference-logprobs/meta.json
reference-logprobs/batch_*.safetensors
The reference was generated from:
zai-org/GLM-5.2
snapshot:… See the full description on the dataset page: https://huggingface.co/datasets/festr2/GLM-5.2-Unsloth-Style-KLD-Refs-20260707.glm51-kld-reference-logits-wikitext-ctx2048-s512-20260517 ---
license: other
pretty_name: GLM-5.1 KLD Reference Logits WikiText ctx2048 s512
tags:
- logits
- kld
- glm-5.1
- vllm
- b12x
---
# GLM-5.1 KLD Reference Logits
Public cache of the reference logits used for GLM-5.1 NVFP4 / mixed
FP8_PB_WO KLD evaluation. These files are generated logits, not model
weights. They are stored as `logits_*.safetensors` with one tensor named
`logits`, shape `(2047, 154880)`, dtype `float32`.
##… See the full description on the dataset page: https://huggingface.co/datasets/festr2/glm51-kld-reference-logits-wikitext-ctx2048-s512-20260517.GLM-5.2-BF16-KLD-Reference-Logits-20260618
GLM-5.2 BF16 KLD Reference Logits
Reference logits for local GLM-5.2 KLD checks.
Contents:
prefill/logits_0.safetensors: BF16 prefill prompt logits generated from
zai-org/GLM-5.2 with context length 2048, stride 512, one window.
decode/decode_teacher_bf16_ref_ctx2048_t17_20260618.safetensors: BF16
teacher-forced decode logits for prompt length 2048 and 17 decode tokens.
decode/decode_teacher_bf16_ref_ctx2048_t17_20260618.safetensors.json:
metadata for the decode reference.… See the full description on the dataset page: https://huggingface.co/datasets/festr2/GLM-5.2-BF16-KLD-Reference-Logits-20260618.open-web-math_TinyLlama_v1.1_KLdiv_Llama-2-7b-hf_TinyLlama_v1.1GLM-5.2-BF16-KLD-Reference-Logits-20260708
GLM-5.2 BF16 KLD Reference Logits 20260708
This dataset contains the current GLM-5.2 BF16 reference prompt logits used for
the July 2026 vLLM/Blackwell KLD checks.
The reference cache is intended for candidate-side KLD comparisons without
rerunning the expensive BF16 reference pass.
Files
reference-logits/logits_0.safetensors
reference-logits/manifest.json
generation-log/config.env
generation-log/scoremode_kld.log
Reference Generation
Field… See the full description on the dataset page: https://huggingface.co/datasets/festr2/GLM-5.2-BF16-KLD-Reference-Logits-20260708.
