Dataset Viewer

The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.

Basalt BF16 KL-divergence bases

The full-precision references that Basalt scores its packs against. With these, anyone can run the same quality gate: PPL, mean KLD and top-1 agreement against the original model in BF16, on their own build or self-converted pack.

File Model Made from Base PPL Size sha256
flashnext-bf16-base.kld Qwen3.8-Flash-Next unsloth's BF16 GGUF through llama-perplexity (llama.cpp 436f6f89e), -b 8192 -ub 8192 3.2749 13GB c4e0221632607324d314556b8813d772867c300454ecdea37f1bf9d5555cf1a9
dsv41-bf16-base.kld DeepSeek-V4.1-Flash a BF16 GGUF by @giveen (convert_hf_to_gguf --outtype bf16, llama.cpp 330f0189) 2.7394 6.6GB f3b1a7e2d6d9ab47b98f7c18de8558a2455f907aae2f26dd53144f484aa63608
qwen38-27b-bf16-base.kld Qwen3.8-27B unsloth's BF16 GGUF through llama-perplexity (llama.cpp 5676dfbe0), -b 8192 -ub 8192 6.7437 13GB 9ffb3bd9ca44bd10b7fbf1f52b6691aad929778704299f084e65fdd19e482f30
hy4flash-bf16-base.kld Hy4-flash Tencent's BF16 checkpoint converted to a BF16 GGUF (llama.cpp 0a054d0b0), through llama-perplexity with the Engram epsilon fix and add_bos_token=true (see below) 18.572 6.2GB abab2c9fcd1026647745befcf79e61c0e9d8376be377a5c90111c8d18910f239

All four bases cover wikitext-2 (wiki.test.raw) in 100 chunks of 512 tokens, written by llama-perplexity --kl-divergence-base. They store the tokens and the full-vocabulary log-probabilities of every scored position.

Use

basalt-bench --model PACK.basalt --quality 100 --kld-base flashnext-bf16-base.kld

Use 20 chunks for a quick check and 100 for a numerics change. The method (scored rows, verify windows, pinned placement for identity) is in How quality is measured. Scores move slightly with placement, so pin the tiers when comparing two builds.

Hy4-flash notes

  • Engram: the llama.cpp port's Engram key/query norm epsilon was patched to match vLLM (BF16 epsilon) before the base was written.
  • BOS: llama-perplexity replaces each chunk's first token with BOS when it evaluates, but writes the original tokens to the file. The token block was corrected so chunks 1 to 99 store BOS (id 120000) at position 0, as evaluated (99 tokens; log-probabilities untouched). Before the fix the file hashed 72f468f9dba50577d7a2b0797c8fa8ce4a575505dfd06001e3e1246e162478f3.
  • The base PPL, 18.572, is over the 25,500 scored rows (256 to 510 of each chunk).

Credits

The DeepSeek base comes from @giveen's BF16 GGUF. The Flash-Next and 27B bases come from unsloth's BF16 GGUFs. The Hy4-flash base comes from Tencent's BF16 checkpoint. All are derived from the original models and follow their licenses.

Downloads last month
482