Modelpublic

eouya2/DeepSeek-V4-Flash-REAP25-REAPDataset10K-BalancedWithKO-DS4

sourceHugging Faceotherupdated 4mo agoView on Hugging Face

2likes47downloads

Model Card

DeepSeek-V4-Flash REAP25 REAPDataset10K-Balanced DS4 GGUF

Experimental DS4 compact GGUF made by applying 25% REAP expert pruning to a DeepSeek-V4-Flash DS4 GGUF, calibrated on 10,000 language-balanced prompts drawn from 8 domains of the REAP dataset.

Model file:

text

DeepSeek-V4-Flash-REAP25-REAPDataset10K-Balanced-DS4-compact-IQ2XXS.gguf

Bundled runtime:

text

ds4_reap_runtime/

Expert observation results:

text

reap_dataset_10k_balanced_seed42_reap25_experts.csv

Compatibility

This model needs the bundled REAP-aware DS4 runtime, or another DS4 build that supports ds4-compact-v1.

It is not expected to run with stock DS4, llama.cpp, Ollama, LM Studio, or other generic GGUF loaders. The routed expert tensors are physically compacted, so the runtime must read the REAP metadata and route into compact expert ids.

Expected DS4 runtime line:

text

REAP runtime metadata enabled: hash_preserved=3 router_masked=40 moe_disabled=0 layout=ds4-compact-v1

How It Was Made

Source GGUF

text

DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix.gguf

Calibration Dataset

Category	Source Dataset	Samples	EN	KO
mixture/code	open-r1/codeforces-cots	2,000	1,000	1,000
mixture/math	open-r1/OpenR1-Math-220k	2,000	1,000	1,000
mixture/science	nvidia/Llama-Nemotron-Post-Training-Dataset	2,000	1,000	1,000
xlam/function-calling	Salesforce/xlam-function-calling-60k	2,000	1,000	1,000
SWE/tool	SWE-bench style (tool-use split)	500	250	250
SWE/xml	SWE-bench style (XML format split)	500	250	250
SWE/ticks	SWE-bench style (tick-format split)	500	250	250
SWE/train	SWE-bench style (training split)	500	250	250
Total		10,000	5,000	5,000

—Sampling: random with seed 42
—Language balance: --balance-language enforced 50% English / 50% Korean per source category
—Total token coverage: 27,592,731 observed prompt tokens
—Observed expert route selections: 7,118,924,598

Observation

—Seed: 42
—Context length: 4,096
—Chunk size: 100 prompts per chunk (100 chunks total, resumable)
—Score metric: activation_energy_sum2

Pruning

—Layers 0–2: preserved, hash-routed
—Layers 3–42: REAP-pruned
—Compression ratio: 0.25
—Experts per pruned layer: 256 → 192 (64 pruned per layer)
—Top-k remains 6
—Layout: ds4-compact-v1
—Expert tensor bytes are copied directly, preserving source quantization

Size

text

source file: 80.76 GiB / 86.72 GB
REAP25 file: 63.87 GiB / 68.58 GB

Local Metal mapping at --ctx 512:

text

source mapped: 82697.67 MiB
REAP25 mapped: 65397.66 MiB
saved: ~17300 MiB, about 16.9 GiB

Expert CSV

reap_dataset_10k_balanced_seed42_reap25_experts.csv contains per-expert statistics for all 43 MoE layers. Columns:

Column	Description
`layer`	Layer index (0–42)
`expert_id`	Original expert ID in source GGUF
`new_expert_id`	Compacted expert ID after pruning (-1 if pruned)
`activation_policy`	`hash_preserved` (layers 0–2) / `router_mask_pruned`
`kept`	Whether this expert is kept in the pruned GGUF
`pruned`	Whether this expert was removed
`total_tokens`	Total observed tokens (shared per layer)
`expert_frequency`	How many times this expert was selected
`selection_rate_per_token`	expertfrequency / totaltokens
`selection_share`	Fraction of all expert selections for this layer
`reap`	Composite REAP score (activationenergysum2)
`gate_up_energy`	Gate/up projection energy contribution
`down_energy`	Down projection energy contribution

Run With Bundled Runtime

The Metal runtime loads shader source files from metal/*.metal, so run from inside the bundled runtime directory:

bash

cd ds4_reap_runtime

./ds4 \
  -m ../DeepSeek-V4-Flash-REAP25-REAPDataset10K-Balanced-DS4-compact-IQ2XXS.gguf \
  --ctx 512 --nothink --temp 0 -n 64 \
  -p 'Hello!'

For OpenAI-compatible local serving:

bash

cd ds4_reap_runtime

./ds4-server \
  -m ../DeepSeek-V4-Flash-REAP25-REAPDataset10K-Balanced-DS4-compact-IQ2XXS.gguf \
  --ctx 32768 --tokens 1024 \
  --host 127.0.0.1 --port 8000

Comparison with LCB50 Model

Property	REAP25-LCB50	REAP25-REAPDataset10K-Balanced (this)
Calibration dataset	LiveCodeBench	REAP dataset (8 domains)
Sample count	50	10,000
Language balance	English only	50% EN / 50% KO
Domain coverage	Competitive coding	Code, Math, Science, Function-calling, SWE
Prompt tokens observed	26,386	27,592,731
Expert route selections	6,807,588	7,118,924,598
Compression	REAP25 (256→192 experts)	REAP25 (256→192 experts)
Output size	63.87 GiB	63.87 GiB

Notes

This is a broader calibration artifact than the LCB50 model. The 10K balanced dataset covers coding, math, science, function-calling, and software engineering domains, with equal Korean and English coverage, providing more representative expert activation statistics.

The REAP pruning removes the 64 least-activated routed experts per layer (layers 3–42) and physically compacts the remaining 192 into a smaller GGUF, so the runtime must read the REAP routing metadata rather than using the original expert slot layout.