Genrius-AI/2bit-gguf-wikitext2-eval

2-bit GGUF WikiText-2 eval: per-window measurements

Raw per-window perplexity measurements behind the 2-bit GGUF leaderboard. Published 2-bit GGUF files of Qwen3.8-27B and Qwen3.5-2B (DASH-Q by jkim96, unsloth UD, and our Pingala-8) were scored on the same text, context and stride, so any two files can be compared window by window.

DASH-Q ranks first on both models.

model	rank	file	bytes	windows	ppl	paired dNLL vs best (95%)	windows won vs best
Qwen3.8-27B	1	DASH-Q IQ2_XS	8,551,168,768	20	6.5431	reference	-
Qwen3.8-27B	2	Pingala-8 v2	8,314,241,600	20	6.8894	+0.0516 +/- 0.0117	0/20
Qwen3.8-27B	3	unsloth UD-IQ2_S	8,371,970,048	20	7.3447	+0.1156 +/- 0.0702	2/20
Qwen3.8-27B	4	Pingala-8 v1	8,314,489,600	20	7.3449	+0.1156 +/- 0.0132	0/20
Qwen3.8-27B	5	llama.cpp IQ2_XXS (local, not published)	8,433,183,520	20	7.8765	+0.1855 +/- 0.0303	0/20
Qwen3.5-2B	1	DASH-Q IQ2_M	786,259,840	72	13.3331	reference	-
Qwen3.5-2B	2	DASH-Q IQ2_XS	757,096,320	72	13.9028	+0.0418 +/- 0.0060	5/72
Qwen3.5-2B	3	Pingala-8 v2	778,189,408	72	14.0573	+0.0529 +/- 0.0081	3/72
Qwen3.5-2B	4	unsloth UD-IQ2_M	859,857,152	72	14.9555	+0.1148 +/- 0.0087	0/72
Qwen3.5-2B	5	Pingala-8 v1	778,744,192	72	15.6193	+0.1583 +/- 0.0107	0/72
Qwen3.5-2B	6	unsloth UD-IQ2_XXS	768,270,592	72	23.4124	+0.5630 +/- 0.0177	0/72

dNLL is in nats per token, file minus best; positive means more loss than the best file.

Files

• windows.csv: one row per (file, window). Columns: model, quant, file, repo, bytes, window, cum_ppl, nll_window, runtime, placement, source_log. cum_ppl is the running perplexity printed by llama-perplexity after window window (4 decimals, as logged). nll_window is the mean NLL of that window in nats per token, derived from cum_ppl (below).
• summary.csv: one row per file: final perplexity, percent above the best file, paired dNLL vs the best file, its 95% interval, and windows won.
• ot/*.ot.txt: the exact -ot regex used per rival file (tensors of type IQ1_S, IQ2_S or IQ3_S, placed on the CPU).

Source files

model	file	repo	bytes
Qwen3.8-27B	Qwen3.8-27B-DASHQ-IQ2_XS.gguf	jkim96/Qwen3.8-27B-DASHQ-Q2-GGUF	8,551,168,768
Qwen3.8-27B	Qwen3.8-27B-UD-IQ2_S.gguf	unsloth/Qwen3.8-27B-GGUF	8,371,970,048
Qwen3.8-27B	Qwen3.8-27B-Pingala8-2bit-0000{1..9}-of-00009.gguf (v2)	Genrius-AI/Qwen3.8-27B-Pingala8-2bit	8,314,241,600
Qwen3.8-27B	same names, v1	revision 81c9a6d8	8,314,489,600
Qwen3.8-27B	q38-27b-IQ2_XXS.gguf (stock llama-quantize IQ2_XXS, same imatrix as Pingala-8)	not published	8,433,183,520
Qwen3.5-2B	Qwen3.5-2B-DASHQ-IQ2_M.gguf, Qwen3.5-2B-DASHQ-IQ2_XS.gguf	jkim96/Qwen3.5-2B-DASHQ-Q2-GGUF	786,259,840; 757,096,320
Qwen3.5-2B	Qwen3.5-2B-UD-IQ2_M.gguf, Qwen3.5-2B-UD-IQ2_XXS.gguf	unsloth/Qwen3.5-2B-GGUF	859,857,152; 768,270,592
Qwen3.5-2B	Qwen3.5-2B-Pingala8-2bit-v2.gguf (v2), Qwen3.5-2B-Pingala8-2bit.gguf (v1)	Genrius-AI/Qwen3.5-2B-Pingala8-2bit	778,189,408; 778,744,192

Byte sizes match the files listed in each repo on 2026-11-01.

Text

WikiText-2 raw test split (Hugging Face wikitext-2-raw-v1 test rows joined), saved as wikitext2_test_hfjoin.raw: 1,296,370 bytes, sha256 696cca6b65a171b0a358a4be6732cdfdf2dd6164a32e20fd70e3c13fc4dfae83, 297,193 tokens with the Qwen tokenizer. Strided evaluation: -c 2049 --ppl-stride 4095 makes llama-perplexity use 4096-token windows with stride 4095, 72 windows in all. Qwen3.8-27B: first 20 windows (--chunks 20, time limit; the v1 and IQ2_XXS runs went to 40 windows and only their first 20 are used). Qwen3.5-2B: all 72 windows.

Exact commands

Common flags: -f wikitext2_test_hfjoin.raw -c 2049 --ppl-stride 4095 -ngl 99 -lm dio -b 512 -ub 512.

Rival files, Qwen3.8-27B (llama.cpp upstream build):

llama-perplexity -m Qwen3.8-27B-DASHQ-IQ2_XS.gguf -f wikitext2_test_hfjoin.raw -c 2049 --ppl-stride 4095 -ngl 99 -lm dio -b 512 -ub 512 \
  --chunks 20 -ot "$(cat ot/Qwen3.8-27B-DASHQ-IQ2_XS.ot.txt)" --no-op-offload
llama-perplexity -m Qwen3.8-27B-UD-IQ2_S.gguf -f wikitext2_test_hfjoin.raw -c 2049 --ppl-stride 4095 -ngl 99 -lm dio -b 512 -ub 512 \
  --chunks 20 -ot "$(cat ot/Qwen3.8-27B-UD-IQ2_S.ot.txt)" --no-op-offload
llama-perplexity -m q38-27b-IQ2_XXS.gguf -f wikitext2_test_hfjoin.raw -c 2049 --ppl-stride 4095 -ngl 99 -lm dio -b 512 -ub 512 --chunks 40

Rival files, Qwen3.5-2B (llama.cpp upstream build, 12 threads; the two IQ2_M files ran with CUDA hidden, so fully on the CPU):

llama-perplexity -m Qwen3.5-2B-<QUANT>.gguf -f wikitext2_test_hfjoin.raw -c 2049 --ppl-stride 4095 -ngl 99 -lm dio -b 512 -ub 512 \
  -t 12 -ot "$(cat ot/Qwen3.5-2B-<QUANT>.ot.txt)" --no-op-offload

Pingala-8 files (Pingala-8 runtime):

GGML_F44=4 GGML_F44_BATCH=2 GGML_VR_NATIVE=1 GGML_VRC_V=5 GGML_VRC_NWP=8 GGML_VR_RL=1 GGML_VR_PRMG=1 GGML_VR_PPC=2 \
llama-perplexity -m Qwen3.8-27B-Pingala8-2bit-00001-of-00009.gguf -f wikitext2_test_hfjoin.raw -c 2049 --ppl-stride 4095 -ngl 99 -lm dio -b 512 -ub 512 --chunks 20
# Qwen3.5-2B: same env and flags, -m Qwen3.5-2B-Pingala8-2bit-v2.gguf (or the v1 file), no --chunks (all 72 windows)
# Qwen3.8-27B v1: same env and flags, --chunks 40 (first 20 used)

Runtime builds

• Rival files and the stock IQ2_XXS reference: llama.cpp upstream commit 889f2ec (2026-10-30), CUDA build for sm_120. One local change to tools/perplexity/perplexity.cpp (strided path only): each batch is scored as soon as its logits are ready instead of keeping all logits for the window, to save host memory. The tokens scored and the formula are unchanged.
• Pingala-8 files: the patched Pingala-8 runtime (a llama.cpp fork with the Pingala-8 CUDA kernels). Pingala-8 numbers are for that runtime.

Hardware

One laptop: NVIDIA RTX 5070 Ti Laptop GPU (12 GB, sm_120), Intel Core Ultra 9 275HX, Windows host with WSL2 (Linux 6.18). On this GPU the upstream IQ1_S/IQ2_S/IQ3_S CUDA kernels failed test-backend-ops, so tensors of those types in rival files ran on the CPU (-ot ...=CPU --no-op-offload). A CPU/GPU cross-check agreed within 0.2%.

Per-window NLL and the paired statistic

llama-perplexity prints the running perplexity [k]P_k after window k. Each window has the same number of scored tokens, so

NLL_1 = ln P_1
NLL_k = k ln P_k - (k-1) ln P_(k-1)

The paired statistic between files A and B is the mean over windows of NLL_A - NLL_B, with a 95% interval of 1.96 * sd / sqrt(n) and the count of windows where A is lower. This is the same derivation as our script r59_paired70.py; recomputed values reproduce the paired lines logged at run time exactly (for example Qwen3.5-2B Pingala-8 v2 minus DASH-Q IQ2_M over 70 windows: +0.0527 +/- 0.0083, 3/70). exp(mean(nll_window)) equals the final cum_ppl of each file to 4 decimals.

Caveats

• Perplexity on one English text; no downstream task scores. The 27B set has 20 windows (about 82k tokens), so its intervals are wider.
• cum_ppl is logged to 4 decimals, so nll_window carries a rounding error that grows with k (at most about 3e-4 nats at k = 72).
• Rival and Pingala-8 files ran on different builds and, for some tensors, different devices (placement column).
• The Qwen3.8-27B IQ2_XXS row is our own stock llama-quantize file, made with the imatrix used for Pingala-8; it is not published.
• Pingala-8 v1 for Qwen3.8-27B is an earlier revision of the same repo, replaced by v2.
• Genrius AI makes Pingala-8. Every window is published so anyone can check the ranking.

License

CC BY 4.0. Measured by Genrius AI. Paper: DOI 10.5281/zenodo.23481757.