GGUF
Chinese
imatrix
conversational

底座模型:gemma-4-26B-A4B-it-heretic
imatrix基于五个中文类数据集,ctx=2048,chunks=512,剔除所有Emoji以及英文单词行,最终得到596万字纯中文作为输入集。

该模型仅对对于泛化性操作(多语言/编写代码等方面),可能不如其他量化版本。


PPL测试结果:

这里拉了一段完整的RP卡作为评测集,89000字左右,用于评测PPL。

Imatrix:

Q5_K_M: PPL = 41.7158 +/- 1.05441

perplexity: 8.55 seconds per pass - ETA 4.55 minutes
[1]42.7054,[2]30.2909,[3]28.7004,[4]28.0079,[5]27.4540,[6]24.8720,[7]29.1118,[8]28.2985,[9]29.4256,[10]31.2614,[11]31.9318,[12]31.9838,[13]31.3226,[14]30.8078,[15]30.7190,[16]31.2388,[17]30.3229,[18]37.5370,[19]37.4290,[20]38.6303,[21]38.3846,[22]37.5415,[23]37.4147,[24]37.5149,[25]36.6674,[26]36.0477,[27]35.8094,[28]37.5164,[29]37.7974,[30]37.8115,[31]41.7254,[32]41.7158,
Final estimate: PPL = 41.7158 +/- 1.05441

Q4_K_M: PPL = 43.6571 +/- 1.09788

perplexity: 8.32 seconds per pass - ETA 4.43 minutes
[1]40.2634,[2]27.5471,[3]27.2934,[4]25.6302,[5]24.7980,[6]22.6113,[7]27.1034,[8]26.4641,[9]27.4364,[10]29.5109,[11]30.6592,[12]32.4097,[13]31.6274,[14]31.3735,[15]31.3907,[16]31.9458,[17]30.8104,[18]37.2838,[19]36.6702,[20]37.8273,[21]37.3376,[22]36.6424,[23]36.0437,[24]36.5134,[25]36.3847,[26]35.9738,[27]35.9467,[28]37.3036,[29]38.0624,[30]37.9556,[31]41.4221,[32]43.6571,
Final estimate: PPL = 43.6571 +/- 1.09788

IQ4_XS: PPL = 48.6481 +/- 1.27893

perplexity: 6.76 seconds per pass - ETA 3.60 minutes
[1]56.8178,[2]37.9150,[3]35.5188,[4]30.8414,[5]30.0177,[6]26.8122,[7]31.0608,[8]30.8852,[9]32.6146,[10]36.0714,[11]37.6788,[12]37.2891,[13]36.8080,[14]36.4927,[15]37.2115,[16]38.4046,[17]36.9525,[18]45.1330,[19]45.2458,[20]46.7740,[21]46.0036,[22]44.8881,[23]44.7756,[24]45.0148,[25]44.3163,[26]43.0405,[27]43.2473,[28]45.3062,[29]45.9349,[30]45.6817,[31]49.1159,[32]48.6481,
Final estimate: PPL = 48.6481 +/- 1.27893

IQ3_XXS: PPL = 39.3126 +/- 0.93952

perplexity: 6.07 seconds per pass - ETA 3.23 minutes
[1]41.0033,[2]26.1371,[3]24.7727,[4]23.5618,[5]23.3471,[6]23.1080,[7]25.4631,[8]25.6110,[9]27.2752,[10]29.1632,[11]29.0638,[12]29.9699,[13]30.4047,[14]30.0000,[15]30.2590,[16]30.7776,[17]30.2628,[18]36.2746,[19]35.6784,[20]36.9668,[21]36.9711,[22]36.2761,[23]35.7942,[24]35.8439,[25]35.7569,[26]35.1833,[27]35.2329,[28]36.4530,[29]37.0048,[30]37.5546,[31]39.7510,[32]39.3126,
Final estimate: PPL = 39.3126 +/- 0.93952

对比组:

gemma-4-26B-A4B-heretic-APEX-I-Quality: PPL = 48.6938 +/- 1.25948

perplexity: 9.30 seconds per pass - ETA 4.95 minutes
[1]40.7008,[2]30.2781,[3]32.6458,[4]31.6194,[5]30.5565,[6]27.7476,[7]31.7446,[8]31.2653,[9]32.7107,[10]35.0757,[11]36.4731,[12]37.8056,[13]36.4391,[14]35.7723,[15]35.4481,[16]36.4941,[17]35.5500,[18]43.2589,[19]42.9500,[20]45.4924,[21]45.2066,[22]43.5416,[23]43.2288,[24]43.5044,[25]42.8942,[26]41.6601,[27]41.4900,[28]43.5188,[29]44.2139,[30]44.2739,[31]48.6861,[32]48.6938,
Final estimate: PPL = 48.6938 +/- 1.25948

全量Q6_K不使用Imatrix: PPL = 47.5879 +/- 1.23391

perplexity: 9.69 seconds per pass - ETA 5.15 minutes
[1]51.4916,[2]35.0155,[3]32.6720,[4]31.6401,[5]30.4067,[6]28.1260,[7]31.7105,[8]31.1109,[9]32.7684,[10]34.8715,[11]35.9251,[12]37.0230,[13]35.4984,[14]34.8886,[15]34.6195,[16]35.7710,[17]34.8576,[18]43.0041,[19]42.7248,[20]44.0729,[21]43.8856,[22]42.6592,[23]42.1386,[24]42.1939,[25]41.5052,[26]40.2828,[27]40.2352,[28]42.3904,[29]43.0217,[30]43.0792,[31]47.8846,[32]47.5879,
Final estimate: PPL = 47.5879 +/- 1.23391
Downloads last month
295
GGUF
Model size
25B params
Architecture
gemma4
Hardware compatibility
Log In to add your hardware

3-bit

4-bit

5-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for wwlwwl/gemma-4-26B-A4B-it-heretic-zh-RP-gguf

Quantized
(24)
this model

Datasets used to train wwlwwl/gemma-4-26B-A4B-it-heretic-zh-RP-gguf