Dataset Viewer
Auto-converted to Parquet Duplicate
name
stringlengths
9
57
dtype
stringclasses
6 values
shape
listlengths
1
3
shard
stringclasses
39 values
begin
int64
664
102B
end
int64
446k
102B
nbytes
int64
12
98.3B
aligner.w1.bias
BF16
[ 5120 ]
model-00001-of-00037.safetensors
1,553,377,760
1,553,388,000
10,240
aligner.w1.weight
BF16
[ 5120, 9216 ]
model-00001-of-00037.safetensors
2,264,745,580
2,359,117,420
94,371,840
aligner.w2.bias
BF16
[ 5120 ]
model-00001-of-00037.safetensors
4,048,921,912
4,048,932,152
10,240
aligner.w2.weight
BF16
[ 5120, 5120 ]
model-00001-of-00037.safetensors
2,206,549,612
2,258,978,412
52,428,800
image_end
BF16
[ 5120 ]
model-00001-of-00037.safetensors
4,356,305,680
4,356,315,920
10,240
image_newline
BF16
[ 5120 ]
model-00001-of-00037.safetensors
1,210,014,916
1,210,025,156
10,240
image_start
BF16
[ 5120 ]
model-00001-of-00037.safetensors
1,955,504,812
1,955,515,052
10,240
language_model.embed.weight
BF16
[ 129280, 5120 ]
model-00001-of-00037.safetensors
2,584,731,352
3,908,558,552
1,323,827,200
language_model.layers.0.attn.attn_sink
F32
[ 64 ]
model-00001-of-00037.safetensors
1,565,553,376
1,565,553,632
256
language_model.layers.0.attn.kv_norm.weight
BF16
[ 512 ]
model-00001-of-00037.safetensors
806,225,596
806,226,620
1,024
language_model.layers.0.attn.q_norm.weight
BF16
[ 1280 ]
model-00001-of-00037.safetensors
2,202,346,508
2,202,349,068
2,560
language_model.layers.0.attn.wkv.scales
U8
[ 512, 160 ]
model-00001-of-00037.safetensors
387,281,152
387,363,072
81,920
language_model.layers.0.attn.wkv.weight
U32
[ 512, 1280 ]
model-00001-of-00037.safetensors
387,363,072
389,984,512
2,621,440
language_model.layers.0.attn.wo_a.weight
BF16
[ 8192, 4096 ]
model-00001-of-00037.safetensors
389,984,512
457,093,376
67,108,864
language_model.layers.0.attn.wo_b.scales
U8
[ 5120, 256 ]
model-00001-of-00037.safetensors
1,429,722,000
1,431,032,720
1,310,720
language_model.layers.0.attn.wo_b.weight
U32
[ 5120, 2048 ]
model-00001-of-00037.safetensors
345,338,112
387,281,152
41,943,040
language_model.layers.0.attn.wq_a.scales
U8
[ 1280, 160 ]
model-00001-of-00037.safetensors
1,431,032,720
1,431,237,520
204,800
language_model.layers.0.attn.wq_a.weight
U32
[ 1280, 1280 ]
model-00001-of-00037.safetensors
4,078,400,824
4,084,954,424
6,553,600
language_model.layers.0.attn.wq_b.scales
U8
[ 32768, 40 ]
model-00001-of-00037.safetensors
457,093,376
458,404,096
1,310,720
language_model.layers.0.attn.wq_b.weight
U32
[ 32768, 320 ]
model-00001-of-00037.safetensors
1,907,281,068
1,949,224,108
41,943,040
language_model.layers.0.attn_norm.weight
BF16
[ 5120 ]
model-00001-of-00037.safetensors
345,327,872
345,338,112
10,240
language_model.layers.0.ffn.gate.bias
F32
[ 384 ]
model-00001-of-00037.safetensors
1,096,373,336
1,096,374,872
1,536
language_model.layers.0.ffn.gate.bias_vl
F32
[ 384 ]
model-00001-of-00037.safetensors
4,037,386,040
4,037,387,576
1,536
language_model.layers.0.ffn.gate.weight
BF16
[ 384, 5120 ]
model-00001-of-00037.safetensors
311,812,352
315,744,512
3,932,160
language_model.layers.0.ffn.shared_experts.w1.scales
U8
[ 2304, 160 ]
model-00001-of-00037.safetensors
344,948,992
345,317,632
368,640
language_model.layers.0.ffn.shared_experts.w1.weight
U32
[ 2304, 1280 ]
model-00001-of-00037.safetensors
990,714,176
1,002,510,656
11,796,480
language_model.layers.0.ffn.shared_experts.w2.scales
U8
[ 5120, 72 ]
model-00001-of-00037.safetensors
5,169,804,236
5,170,172,876
368,640
language_model.layers.0.ffn.shared_experts.w2.weight
U32
[ 5120, 576 ]
model-00001-of-00037.safetensors
315,744,512
327,540,992
11,796,480
language_model.layers.0.ffn.shared_experts.w3.scales
U8
[ 2304, 160 ]
model-00001-of-00037.safetensors
327,540,992
327,909,632
368,640
language_model.layers.0.ffn.shared_experts.w3.weight
U32
[ 2304, 1280 ]
model-00001-of-00037.safetensors
333,152,512
344,948,992
11,796,480
language_model.layers.0.ffn_norm.weight
BF16
[ 5120 ]
model-00001-of-00037.safetensors
345,317,632
345,327,872
10,240
language_model.layers.0.hc_attn_base
F32
[ 24 ]
model-00001-of-00037.safetensors
460,370,188
460,370,284
96
language_model.layers.0.hc_attn_fn
F32
[ 24, 20480 ]
model-00001-of-00037.safetensors
852,184,764
854,150,844
1,966,080
language_model.layers.0.hc_attn_scale
F32
[ 3 ]
model-00001-of-00037.safetensors
921,259,708
921,259,720
12
language_model.layers.0.hc_ffn_base
F32
[ 24 ]
model-00001-of-00037.safetensors
5,241,738,284
5,241,738,380
96
language_model.layers.0.hc_ffn_fn
F32
[ 24, 20480 ]
model-00001-of-00037.safetensors
458,404,108
460,370,188
1,966,080
language_model.layers.0.hc_ffn_scale
F32
[ 3 ]
model-00001-of-00037.safetensors
458,404,096
458,404,108
12
language_model.layers.14.attn.attn_sink
F32
[ 64 ]
model-00001-of-00037.safetensors
1,523,693,660
1,523,693,916
256
language_model.layers.14.attn.compressor.norm.weight
BF16
[ 512 ]
model-00001-of-00037.safetensors
2,201,034,412
2,201,035,436
1,024
language_model.layers.14.attn.compressor.wgate.weight
BF16
[ 512, 5120 ]
model-00001-of-00037.safetensors
4,483,878,964
4,489,121,844
5,242,880
language_model.layers.14.attn.compressor.wkv.weight
BF16
[ 512, 5120 ]
model-00001-of-00037.safetensors
5,033,023,692
5,038,266,572
5,242,880
language_model.layers.14.attn.indexer.k_norm.weight
BF16
[ 128 ]
model-00001-of-00037.safetensors
1,565,553,120
1,565,553,376
256
language_model.layers.14.attn.indexer.weights_proj.weight
BF16
[ 32, 5120 ]
model-00001-of-00037.safetensors
5,350,675,704
5,351,003,384
327,680
language_model.layers.14.attn.indexer.wk.weight
BF16
[ 128, 512 ]
model-00001-of-00037.safetensors
2,142,502,572
2,142,633,644
131,072
language_model.layers.14.attn.indexer.wq_b.scales
U8
[ 4096, 40 ]
model-00001-of-00037.safetensors
4,784,744,608
4,784,908,448
163,840
language_model.layers.14.attn.indexer.wq_b.weight
U32
[ 4096, 320 ]
model-00001-of-00037.safetensors
1,536,231,892
1,541,474,772
5,242,880
language_model.layers.14.attn.kv_norm.weight
BF16
[ 512 ]
model-00001-of-00037.safetensors
2,455,348,332
2,455,349,356
1,024
language_model.layers.14.attn.q_norm.weight
BF16
[ 1280 ]
model-00001-of-00037.safetensors
4,784,908,448
4,784,911,008
2,560
language_model.layers.14.attn.wkv.scales
U8
[ 512, 160 ]
model-00001-of-00037.safetensors
1,567,519,712
1,567,601,632
81,920
language_model.layers.14.attn.wkv.weight
U32
[ 512, 1280 ]
model-00001-of-00037.safetensors
1,567,601,632
1,570,223,072
2,621,440
language_model.layers.14.attn.wo_a.weight
BF16
[ 8192, 4096 ]
model-00001-of-00037.safetensors
1,570,223,072
1,637,331,936
67,108,864
language_model.layers.14.attn.wo_b.scales
U8
[ 5120, 256 ]
model-00001-of-00037.safetensors
4,354,994,960
4,356,305,680
1,310,720
language_model.layers.14.attn.wo_b.weight
U32
[ 5120, 2048 ]
model-00001-of-00037.safetensors
4,091,247,928
4,133,190,968
41,943,040
language_model.layers.14.attn.wq_a.scales
U8
[ 1280, 160 ]
model-00001-of-00037.safetensors
2,033,238,188
2,033,442,988
204,800
language_model.layers.14.attn.wq_a.weight
U32
[ 1280, 1280 ]
model-00001-of-00037.safetensors
4,908,336,384
4,914,889,984
6,553,600
language_model.layers.14.attn.wq_b.scales
U8
[ 32768, 40 ]
model-00001-of-00037.safetensors
1,639,298,016
1,640,608,736
1,310,720
language_model.layers.14.attn.wq_b.weight
U32
[ 32768, 320 ]
model-00001-of-00037.safetensors
1,646,382,048
1,688,325,088
41,943,040
language_model.layers.14.attn_norm.weight
BF16
[ 5120 ]
model-00001-of-00037.safetensors
3,931,895,512
3,931,905,752
10,240
language_model.layers.14.engram.k_weight
BF16
[ 4, 5120 ]
model-00001-of-00037.safetensors
4,420,387,368
4,420,428,328
40,960
language_model.layers.14.engram.q_weight
BF16
[ 4, 5120 ]
model-00001-of-00037.safetensors
4,348,658,448
4,348,699,408
40,960
language_model.layers.14.engram.wkv.scales
U8
[ 25600, 192 ]
model-00001-of-00037.safetensors
1,745,079,372
1,749,994,572
4,915,200
language_model.layers.14.engram.wkv.weight
U32
[ 25600, 1536 ]
model-00001-of-00037.safetensors
1,749,994,668
1,907,281,068
157,286,400
language_model.layers.14.ffn.gate.bias
F32
[ 384 ]
model-00001-of-00037.safetensors
4,406,748,200
4,406,749,736
1,536
language_model.layers.14.ffn.gate.bias_vl
F32
[ 384 ]
model-00001-of-00037.safetensors
5,247,507,692
5,247,509,228
1,536
language_model.layers.14.ffn.gate.weight
BF16
[ 384, 5120 ]
model-00001-of-00037.safetensors
1,442,832,892
1,446,765,052
3,932,160
language_model.layers.14.ffn.shared_experts.w1.scales
U8
[ 2304, 160 ]
model-00001-of-00037.safetensors
1,553,388,000
1,553,756,640
368,640
language_model.layers.14.ffn.shared_experts.w1.weight
U32
[ 2304, 1280 ]
model-00001-of-00037.safetensors
1,553,756,640
1,565,553,120
11,796,480
language_model.layers.14.ffn.shared_experts.w2.scales
U8
[ 5120, 72 ]
model-00001-of-00037.safetensors
1,523,698,132
1,524,066,772
368,640
language_model.layers.14.ffn.shared_experts.w2.weight
U32
[ 5120, 576 ]
model-00001-of-00037.safetensors
4,565,690,932
4,577,487,412
11,796,480
language_model.layers.14.ffn.shared_experts.w3.scales
U8
[ 2304, 160 ]
model-00001-of-00037.safetensors
1,524,066,772
1,524,435,412
368,640
language_model.layers.14.ffn.shared_experts.w3.weight
U32
[ 2304, 1280 ]
model-00001-of-00037.safetensors
1,524,435,412
1,536,231,892
11,796,480
language_model.layers.14.ffn_norm.weight
BF16
[ 5120 ]
model-00001-of-00037.safetensors
4,893,246,624
4,893,256,864
10,240
language_model.layers.14.hc_attn_base
F32
[ 24 ]
model-00001-of-00037.safetensors
1,738,525,676
1,738,525,772
96
language_model.layers.14.hc_attn_fn
F32
[ 24, 20480 ]
model-00001-of-00037.safetensors
1,736,559,596
1,738,525,676
1,966,080
language_model.layers.14.hc_attn_scale
F32
[ 3 ]
model-00001-of-00037.safetensors
5,286,437,612
5,286,437,624
12
language_model.layers.14.hc_ffn_base
F32
[ 24 ]
model-00001-of-00037.safetensors
1,749,994,572
1,749,994,668
96
language_model.layers.14.hc_ffn_fn
F32
[ 24, 20480 ]
model-00001-of-00037.safetensors
4,149,181,752
4,151,147,832
1,966,080
language_model.layers.14.hc_ffn_scale
F32
[ 3 ]
model-00001-of-00037.safetensors
1,688,325,088
1,688,325,100
12
language_model.layers.15.attn.attn_sink
F32
[ 64 ]
model-00001-of-00037.safetensors
990,713,920
990,714,176
256
language_model.layers.15.attn.kv_norm.weight
BF16
[ 512 ]
model-00001-of-00037.safetensors
948,687,936
948,688,960
1,024
language_model.layers.15.attn.q_norm.weight
BF16
[ 1280 ]
model-00001-of-00037.safetensors
4,165,567,812
4,165,570,372
2,560
language_model.layers.15.attn.wkv.scales
U8
[ 512, 160 ]
model-00001-of-00037.safetensors
948,688,960
948,770,880
81,920
language_model.layers.15.attn.wkv.weight
U32
[ 512, 1280 ]
model-00001-of-00037.safetensors
2,570,571,384
2,573,192,824
2,621,440
language_model.layers.15.attn.wo_a.weight
BF16
[ 8192, 4096 ]
model-00001-of-00037.safetensors
2,388,221,548
2,455,330,412
67,108,864
language_model.layers.15.attn.wo_b.scales
U8
[ 5120, 256 ]
model-00001-of-00037.safetensors
2,201,035,436
2,202,346,156
1,310,720
language_model.layers.15.attn.wo_b.weight
U32
[ 5120, 2048 ]
model-00001-of-00037.safetensors
4,845,526,176
4,887,469,216
41,943,040
language_model.layers.15.attn.wq_a.scales
U8
[ 1280, 160 ]
model-00001-of-00037.safetensors
5,100,377,036
5,100,581,836
204,800
language_model.layers.15.attn.wq_a.weight
U32
[ 1280, 1280 ]
model-00001-of-00037.safetensors
4,233,478,052
4,240,031,652
6,553,600
language_model.layers.15.attn.wq_b.scales
U8
[ 32768, 40 ]
model-00001-of-00037.safetensors
4,552,374,836
4,553,685,556
1,310,720
language_model.layers.15.attn.wq_b.weight
U32
[ 32768, 320 ]
model-00001-of-00037.safetensors
948,770,880
990,713,920
41,943,040
language_model.layers.15.attn_norm.weight
BF16
[ 5120 ]
model-00001-of-00037.safetensors
4,374,821,660
4,374,831,900
10,240
language_model.layers.15.ffn.gate.bias
F32
[ 384 ]
model-00001-of-00037.safetensors
5,298,234,104
5,298,235,640
1,536
language_model.layers.15.ffn.gate.bias_vl
F32
[ 384 ]
model-00001-of-00037.safetensors
924,704,308
924,705,844
1,536
language_model.layers.15.ffn.gate.weight
BF16
[ 384, 5120 ]
model-00001-of-00037.safetensors
3,910,657,752
3,914,589,912
3,932,160
language_model.layers.15.ffn.shared_experts.w1.scales
U8
[ 2304, 160 ]
model-00001-of-00037.safetensors
1,431,320,464
1,431,689,104
368,640
language_model.layers.15.ffn.shared_experts.w1.weight
U32
[ 2304, 1280 ]
model-00001-of-00037.safetensors
4,500,931,124
4,512,727,604
11,796,480
language_model.layers.15.ffn.shared_experts.w2.scales
U8
[ 5120, 72 ]
model-00001-of-00037.safetensors
4,838,861,472
4,839,230,112
368,640
language_model.layers.15.ffn.shared_experts.w2.weight
U32
[ 5120, 576 ]
model-00001-of-00037.safetensors
924,705,844
936,502,324
11,796,480
language_model.layers.15.ffn.shared_experts.w3.scales
U8
[ 2304, 160 ]
model-00001-of-00037.safetensors
936,512,564
936,881,204
368,640
language_model.layers.15.ffn.shared_experts.w3.weight
U32
[ 2304, 1280 ]
model-00001-of-00037.safetensors
936,881,216
948,677,696
11,796,480
End of preview. Expand in Data Studio

DeepSeek-V4.1-Flash-1017 — dismantle record

Everything recoverable from the shipped MLX pack of deepseek-ai/DeepSeek-V4.1-Flash-1017 without the weights themselves: the full architecture, the exact engram hash (verified bit-exact against the checkpoint's own row counts), the quantization codecs, per-tensor shapes and byte offsets, router statistics for all 40 layers, the draft-stage (DSpark) structure, the chat template, and a deterministic capability battery with teacher outputs.

Produced by dismantling ddalcu/DeepSeek-V4.1-Flash-1017-MLX-Serve-iQ-MLX-3.0bpw (MIT) — 1,964 tensors, 427.288 GB, 39 shards.

What is here

path what it holds
spec/ARCHITECTURE.md every layer's shape, sparse-attention pattern, hyper-connections, vision, draft stages
spec/ENGRAM.md geometry, hash, token map, 264-byte row format, gating
spec/engram_reference.py runnable hash + token-map unpack; sums to the declared rows exactly
spec/engram_tables.json the 48 per-layer hash constants (multipliers, primes, offsets)
spec/QUANTIZATION.md affine g64 2/3-bit, MXFP8, MXFP4; where the bytes go
spec/quant_codec.py dequantizers for all four modes
spec/config.json the release's text_config + vision_config + omlx_deepseek_v41
spec/model.safetensors.index.json the shipped weight map
spec/tensor_manifest.json all 1,964 tensors: dtype, shape, shard, byte begin/end (also data/tensors.jsonl)
spec/dsv41_chat_template.jinja the DSML tool-call / reasoning-effort template
spec/dsv41_engram_token_map.bin packed 129,280 → 99,092 token map
tokenizer/ tokenizer.json (129,280 ids incl. |DSML|, thinking, </think>)
stats/ router usage per layer (calibration-free), gate/bias statistics, training signatures
eval/battery.jsonl 14 deterministic cases with teacher outputs and the exact checkers
eval/copy_probe.py the verbatim-copy probe that exposes engram damage
validation/ four independent validator reports, including the falsified claims
PLAN.md findings summary, compressibility probe, the engram-transplant distillation plan, what is still missing
experts/ real routing capture on the stock pack: 40×384 expert catalogue with per-domain specialisation, ESFT sets, fine-tune shortlists (experts/ANALYSIS.md)
stats/compressibility_probe.json engram dead-row / rank / cosine probe and expert similarity / SVD probe
experts/expert_catalog.csv every layer-expert: share/lift per domain, gate bias, top tokens, ISTA-coder membership
experts/recommendations.json per-domain expert-specialised fine-tuning sets (ESFT), with validation
experts/ABLATION.md measured: what the identified coding experts actually carry (≈nothing alone), pruning cost, and focused-model speed
experts/ablation/ every battery/speed result per arm, with server logs
scripts/ builders used to produce the pruned/stripped variants
provenance/ upstream card, MIT license, SHA-256 sums

Quick use

import json, sys
sys.path.insert(0, "spec")
import engram_reference as E

cfg = json.load(open("spec/config.json"))["text_config"]
tables = E.magic(cfg)                      # verifies buckets == engram_num_embeddings
tm = E.load_token_map(open("spec/dsv41_engram_token_map.bin", "rb").read())
rows = E.ngram_rows(tables, 0, toks=tm[-4:], pad=2)
import json
m = json.load(open("spec/tensor_manifest.json"))
expert_w1 = [t for t in m["tensors"] if t["shard"] == "model-00001-of-00037.safetensors"][:3]

Verification

  • engram_reference.py reproduces the bucket sums the engine enforces at load: layer 1 → 384,006,168; layer 14 → 384,016,682. Both exact.
  • Token map unpacks to 129,280 ids with max + 1 == 99,092.
  • Every quantized tensor's stored shape and scale count match the declared bit-width and group size (checked across all 1,964 tensors).
  • The 14-case battery below is deterministic: the teacher scores 14/14, and each case carries the exact pass predicate, so it can be re-run on any reimplementation.

What this is not

The weights, the pretraining corpus, the optimizer state, the data mixture and the RL data are not here and are not recoverable from an artifact. See REBUILD.md — this dataset makes the model reimplementable, not retrainable to parity. For that you need the numbers in that file, which are deliberately plain.

License

MIT, inherited from the base model and the MLX pack. spec/config.json, tokenizer/ and provenance/UPSTREAM_CARD.md are DeepSeek's / the pack author's, redistributed under the same MIT terms. spec/dsv41_chat_template.jinja and the engram algorithm were recovered from mlx-serve (MIT, © David Dalcu), itself a transcription of DeepSeek's encoding/encoding.py and inference/engram.py.

Downloads last month
150