Dataset Viewer
Auto-converted to Parquet Duplicate
schema_version
string
rescore_config_sha256
string
source_dataset_index_sha256
string
source_experiment_sha256
string
source_parquet_sha256
string
source_generation_config_sha256
string
trace_id
string
direction
string
split
string
source_dataset
string
source_dataset_sha256
string
source_uid
string
question_sha256
string
prompt_index
int64
sample_index
int8
question_text
large_string
gold_answer
large_string
strict_grade
float32
strict_correct
bool
strict_prediction
large_string
response_text
large_string
vllm_response_text
large_string
prompt_token_ids
list
response_token_ids
list
input_ids
list
response_mask
list
teacher_topk_token_ids
list
teacher_topk_logprobs
list
teacher_sampled_token_logprobs
list
teacher_topk_rank_order
string
prompt_length
int32
response_length
int32
shard_id
int32
row_within_shard
int32
source_target_engine
string
source_teacher_model
string
source_teacher_revision
string
source_teacher_content_sha256
string
source_sampling_parameters_json
large_string
source_environment_versions_json
large_string
target_engine
string
target_model_identity_sha256
string
target_dtype
string
target_attention_implementation
string
target_final_logit_softcapping
float32
rescoring_timestamp
string
gemma4-hf-bf16-sdpa-topk-overlay-v1
1c409d348a85f8d66941e7cb0700279af127147d82c39e6079aab929fcfca17a
efe76f5a53225e97081a750495ceb8ebe26b1a72a3ea8e0b74ea079a87828c0a
17ba236cfe2fdc25ed6f2d5e145534abfdd890a5d42396e8b801012301a52f64
a9c4d270f0246252b6e4d3478f99af61b7f41e78e552d76a43025a0fca55bd78
a2c1aacb4166472a6da7c0af223270b0feb289c1d612b08ea5c090184687d75c
756b89501b704c571938a9e377dacf081aca51f7fab0cdf91e7bb070cb834f2d
e2b_base_to_e4b
train
deepscaler_4of4strict_rl_train
407211b033fb4013e7199012f441301e4012caf9d493a1321587bdae3c894918
row:00000000:sha256:2171ec331e6721791c443eda8b396b9dbbf5498659b84df5e56ef56827eda9dc
2171ec331e6721791c443eda8b396b9dbbf5498659b84df5e56ef56827eda9dc
0
0
"Let $a,$ $b,$ $c,$ $d,$ $e,$ $f$ be positive real numbers such that $a + b + c + d + e + f = 7.$ F(...TRUNCATED)
63
0
false
\frac{3}{5}
"We know that $a + b + c + d + e + f = 7$, giving $c = 7 - a - b - d - e - f$. Separating out the nu(...TRUNCATED)
"We know that $a + b + c + d + e + f = 7$, giving $c = 7 - a - b - d - e - f$. Separating out the nu(...TRUNCATED)
[2,236820,3041,236779,1340,236779,887,236813,2364,107,11823,506,8610,529,506,5619,874,1420,1564,4784(...TRUNCATED)
[1882,1281,600,609,236746,900,518,900,505,900,513,900,545,900,517,578,236743,236832,2500,6295,609,23(...TRUNCATED)
[2,236820,3041,236779,1340,236779,887,236813,2364,107,11823,506,8610,529,506,5619,874,1420,1564,4784(...TRUNCATED)
[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0(...TRUNCATED)
[[1882,818,10081,2318,236795,6481,2292,17123,3292,194234,7280,2859,2094,236785,31836,51613,7795,2542(...TRUNCATED)
[[-1.3115234375,-2.185546875,-2.435546875,-2.810546875,-2.935546875,-3.310546875,-3.560546875,-3.560(...TRUNCATED)
[-1.3115234375,-2.5859375,-0.92822265625,-0.297119140625,-0.148193359375,-0.1783447265625,-0.0145034(...TRUNCATED)
1..128
1,633
465
0
0
vllm
/tmp/hf_cache/models--google--gemma-4-E2B/snapshots/d29ff6b45f081a49ee2733a859c9c9c2d95d1a6f
null
76dc84a5a805a2c8b91e9ccc00b8dbf8f4a99bf0d56ab25832f6e6addd4f7f57
"{\"include_stop_str_in_output\":false,\"logprobs\":128,\"max_model_len\":12288,\"max_prompt_tokens\(...TRUNCATED)
"{\"platform\":\"Linux-6.8.0-1046-nvidia-x86_64-with-glibc2.35\",\"pyarrow\":\"25.0.0\",\"python\":\(...TRUNCATED)
hf_bf16_sdpa_full_forward
bde9e800223cdd62228ce39e0305398f6ada05b98adaf438b0b3d3d3c3015561
bfloat16
sdpa
30
2026-07-31T06:26:16.720095+00:00
gemma4-hf-bf16-sdpa-topk-overlay-v1
1c409d348a85f8d66941e7cb0700279af127147d82c39e6079aab929fcfca17a
efe76f5a53225e97081a750495ceb8ebe26b1a72a3ea8e0b74ea079a87828c0a
17ba236cfe2fdc25ed6f2d5e145534abfdd890a5d42396e8b801012301a52f64
a9c4d270f0246252b6e4d3478f99af61b7f41e78e552d76a43025a0fca55bd78
a2c1aacb4166472a6da7c0af223270b0feb289c1d612b08ea5c090184687d75c
45384ba9371a76326c755361de2c8599b799d786da4ec3eca86b60d76ae7c1a4
e2b_base_to_e4b
train
deepscaler_4of4strict_rl_train
407211b033fb4013e7199012f441301e4012caf9d493a1321587bdae3c894918
row:00000000:sha256:2171ec331e6721791c443eda8b396b9dbbf5498659b84df5e56ef56827eda9dc
2171ec331e6721791c443eda8b396b9dbbf5498659b84df5e56ef56827eda9dc
0
1
"Let $a,$ $b,$ $c,$ $d,$ $e,$ $f$ be positive real numbers such that $a + b + c + d + e + f = 7.$ F(...TRUNCATED)
63
0
false
\frac{29}{b}
"Note that we need to set the $a + b + c + d + e + f$ (always positive) to 7. So $a + b + c + d + e (...TRUNCATED)
"Note that we need to set the $a + b + c + d + e + f$ (always positive) to 7. So $a + b + c + d + e (...TRUNCATED)
[2,236820,3041,236779,1340,236779,887,236813,2364,107,11823,506,8610,529,506,5619,874,1420,1564,4784(...TRUNCATED)
[10282,600,692,1202,531,1076,506,609,236746,900,518,900,505,900,513,900,545,900,517,236795,568,47236(...TRUNCATED)
[2,236820,3041,236779,1340,236779,887,236813,2364,107,11823,506,8610,529,506,5619,874,1420,1564,4784(...TRUNCATED)
[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0(...TRUNCATED)
[[1882,818,10081,2318,236795,6481,2292,17123,3292,194234,7280,2859,2094,236785,31836,51613,7795,2542(...TRUNCATED)
[[-1.3115234375,-2.185546875,-2.435546875,-2.810546875,-2.935546875,-3.310546875,-3.560546875,-3.560(...TRUNCATED)
[-5.5625,-0.08258056640625,-3.060546875,-3.669921875,-0.828125,-6.12109375,-2.1484375,-3.716796875,-(...TRUNCATED)
1..128
1,633
262
0
1
vllm
/tmp/hf_cache/models--google--gemma-4-E2B/snapshots/d29ff6b45f081a49ee2733a859c9c9c2d95d1a6f
null
76dc84a5a805a2c8b91e9ccc00b8dbf8f4a99bf0d56ab25832f6e6addd4f7f57
"{\"include_stop_str_in_output\":false,\"logprobs\":128,\"max_model_len\":12288,\"max_prompt_tokens\(...TRUNCATED)
"{\"platform\":\"Linux-6.8.0-1046-nvidia-x86_64-with-glibc2.35\",\"pyarrow\":\"25.0.0\",\"python\":\(...TRUNCATED)
hf_bf16_sdpa_full_forward
bde9e800223cdd62228ce39e0305398f6ada05b98adaf438b0b3d3d3c3015561
bfloat16
sdpa
30
2026-07-31T06:26:16.720095+00:00
gemma4-hf-bf16-sdpa-topk-overlay-v1
1c409d348a85f8d66941e7cb0700279af127147d82c39e6079aab929fcfca17a
efe76f5a53225e97081a750495ceb8ebe26b1a72a3ea8e0b74ea079a87828c0a
17ba236cfe2fdc25ed6f2d5e145534abfdd890a5d42396e8b801012301a52f64
a9c4d270f0246252b6e4d3478f99af61b7f41e78e552d76a43025a0fca55bd78
a2c1aacb4166472a6da7c0af223270b0feb289c1d612b08ea5c090184687d75c
0fa0326cb96f76edb276c42a80b0361926f8e4151f90d0aac88e7d47a5081e5b
e2b_base_to_e4b
train
deepscaler_4of4strict_rl_train
407211b033fb4013e7199012f441301e4012caf9d493a1321587bdae3c894918
row:00000000:sha256:2171ec331e6721791c443eda8b396b9dbbf5498659b84df5e56ef56827eda9dc
2171ec331e6721791c443eda8b396b9dbbf5498659b84df5e56ef56827eda9dc
0
2
"Let $a,$ $b,$ $c,$ $d,$ $e,$ $f$ be positive real numbers such that $a + b + c + d + e + f = 7.$ F(...TRUNCATED)
63
0
false
2
"We have $a + b + c + d + e + f = 7.$ Multiplying by $2$ on the left and by $7$ on the right gives $(...TRUNCATED)
"We have $a + b + c + d + e + f = 7.$ Multiplying by $2$ on the left and by $7$ on the right gives $(...TRUNCATED)
[2,236820,3041,236779,1340,236779,887,236813,2364,107,11823,506,8610,529,506,5619,874,1420,1564,4784(...TRUNCATED)
[1882,735,609,236746,900,518,900,505,900,513,900,545,900,517,578,236743,236832,9222,115142,4271,684,(...TRUNCATED)
[2,236820,3041,236779,1340,236779,887,236813,2364,107,11823,506,8610,529,506,5619,874,1420,1564,4784(...TRUNCATED)
[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0(...TRUNCATED)
[[1882,818,10081,2318,236795,6481,2292,17123,3292,194234,7280,2859,2094,236785,31836,51613,7795,2542(...TRUNCATED)
[[-1.3115234375,-2.185546875,-2.435546875,-2.810546875,-2.935546875,-3.310546875,-3.560546875,-3.560(...TRUNCATED)
[-1.3115234375,-0.8359375,-0.6630859375,-0.30908203125,-0.10491943359375,-0.026214599609375,-0.00719(...TRUNCATED)
1..128
1,633
146
0
2
vllm
/tmp/hf_cache/models--google--gemma-4-E2B/snapshots/d29ff6b45f081a49ee2733a859c9c9c2d95d1a6f
null
76dc84a5a805a2c8b91e9ccc00b8dbf8f4a99bf0d56ab25832f6e6addd4f7f57
"{\"include_stop_str_in_output\":false,\"logprobs\":128,\"max_model_len\":12288,\"max_prompt_tokens\(...TRUNCATED)
"{\"platform\":\"Linux-6.8.0-1046-nvidia-x86_64-with-glibc2.35\",\"pyarrow\":\"25.0.0\",\"python\":\(...TRUNCATED)
hf_bf16_sdpa_full_forward
bde9e800223cdd62228ce39e0305398f6ada05b98adaf438b0b3d3d3c3015561
bfloat16
sdpa
30
2026-07-31T06:26:16.720095+00:00
gemma4-hf-bf16-sdpa-topk-overlay-v1
1c409d348a85f8d66941e7cb0700279af127147d82c39e6079aab929fcfca17a
efe76f5a53225e97081a750495ceb8ebe26b1a72a3ea8e0b74ea079a87828c0a
17ba236cfe2fdc25ed6f2d5e145534abfdd890a5d42396e8b801012301a52f64
a9c4d270f0246252b6e4d3478f99af61b7f41e78e552d76a43025a0fca55bd78
a2c1aacb4166472a6da7c0af223270b0feb289c1d612b08ea5c090184687d75c
c18100ccd7611ed872df626b9139c7a8837db5c30b2dfa8bc659f79496471f45
e2b_base_to_e4b
train
deepscaler_4of4strict_rl_train
407211b033fb4013e7199012f441301e4012caf9d493a1321587bdae3c894918
row:00000000:sha256:2171ec331e6721791c443eda8b396b9dbbf5498659b84df5e56ef56827eda9dc
2171ec331e6721791c443eda8b396b9dbbf5498659b84df5e56ef56827eda9dc
0
3
"Let $a,$ $b,$ $c,$ $d,$ $e,$ $f$ be positive real numbers such that $a + b + c + d + e + f = 7.$ F(...TRUNCATED)
63
0
false
35
"All terms are from 0 to 7, so only one term can be $1$, so this is equivalent to finding the minimu(...TRUNCATED)
"All terms are from 0 to 7, so only one term can be $1$, so this is equivalent to finding the minimu(...TRUNCATED)
[2,236820,3041,236779,1340,236779,887,236813,2364,107,11823,506,8610,529,506,5619,874,1420,1564,4784(...TRUNCATED)
[3243,3755,659,699,236743,236771,531,236743,236832,236764,834,1186,886,1941,740,577,609,236770,2500,(...TRUNCATED)
[2,236820,3041,236779,1340,236779,887,236813,2364,107,11823,506,8610,529,506,5619,874,1420,1564,4784(...TRUNCATED)
[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0(...TRUNCATED)
[[1882,818,10081,2318,236795,6481,2292,17123,3292,194234,7280,2859,2094,236785,31836,51613,7795,2542(...TRUNCATED)
[[-1.3115234375,-2.185546875,-2.435546875,-2.810546875,-2.935546875,-3.310546875,-3.560546875,-3.560(...TRUNCATED)
[-6.0625,-2.091796875,-0.63037109375,-6.83203125,-1.0419921875,-1.1552734375,-0.06439208984375,-0.10(...TRUNCATED)
1..128
1,633
202
0
3
vllm
/tmp/hf_cache/models--google--gemma-4-E2B/snapshots/d29ff6b45f081a49ee2733a859c9c9c2d95d1a6f
null
76dc84a5a805a2c8b91e9ccc00b8dbf8f4a99bf0d56ab25832f6e6addd4f7f57
"{\"include_stop_str_in_output\":false,\"logprobs\":128,\"max_model_len\":12288,\"max_prompt_tokens\(...TRUNCATED)
"{\"platform\":\"Linux-6.8.0-1046-nvidia-x86_64-with-glibc2.35\",\"pyarrow\":\"25.0.0\",\"python\":\(...TRUNCATED)
hf_bf16_sdpa_full_forward
bde9e800223cdd62228ce39e0305398f6ada05b98adaf438b0b3d3d3c3015561
bfloat16
sdpa
30
2026-07-31T06:26:16.720095+00:00
gemma4-hf-bf16-sdpa-topk-overlay-v1
1c409d348a85f8d66941e7cb0700279af127147d82c39e6079aab929fcfca17a
efe76f5a53225e97081a750495ceb8ebe26b1a72a3ea8e0b74ea079a87828c0a
17ba236cfe2fdc25ed6f2d5e145534abfdd890a5d42396e8b801012301a52f64
a9c4d270f0246252b6e4d3478f99af61b7f41e78e552d76a43025a0fca55bd78
a2c1aacb4166472a6da7c0af223270b0feb289c1d612b08ea5c090184687d75c
d6cea30b77412c277bdf71300622672a5b19e3c59f4d1aac24def80cfc3419af
e2b_base_to_e4b
train
deepscaler_4of4strict_rl_train
407211b033fb4013e7199012f441301e4012caf9d493a1321587bdae3c894918
row:00000000:sha256:2171ec331e6721791c443eda8b396b9dbbf5498659b84df5e56ef56827eda9dc
2171ec331e6721791c443eda8b396b9dbbf5498659b84df5e56ef56827eda9dc
0
4
"Let $a,$ $b,$ $c,$ $d,$ $e,$ $f$ be positive real numbers such that $a + b + c + d + e + f = 7.$ F(...TRUNCATED)
63
0
false
36/77
"By the Law of Small Integer (for positive integers $n, m, a, b, c > 0$), we have $$-\\frac{1}{a} + (...TRUNCATED)
"By the Law of Small Integer (for positive integers $n, m, a, b, c > 0$), we have $$-\\frac{1}{a} + (...TRUNCATED)
[2,236820,3041,236779,1340,236779,887,236813,2364,107,11823,506,8610,529,506,5619,874,1420,1564,4784(...TRUNCATED)
[2292,506,6563,529,16647,15556,568,1708,4414,25630,609,236749,236764,520,236764,496,236764,518,23676(...TRUNCATED)
[2,236820,3041,236779,1340,236779,887,236813,2364,107,11823,506,8610,529,506,5619,874,1420,1564,4784(...TRUNCATED)
[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0(...TRUNCATED)
[[1882,818,10081,2318,236795,6481,2292,17123,3292,194234,7280,2859,2094,236785,31836,51613,7795,2542(...TRUNCATED)
[[-1.3115234375,-2.185546875,-2.435546875,-2.810546875,-2.935546875,-3.310546875,-3.560546875,-3.560(...TRUNCATED)
[-3.560546875,-0.70751953125,-5.22265625,-0.00461578369140625,-3.138671875,-3.96875,-5.44921875,-5.4(...TRUNCATED)
1..128
1,633
509
0
4
vllm
/tmp/hf_cache/models--google--gemma-4-E2B/snapshots/d29ff6b45f081a49ee2733a859c9c9c2d95d1a6f
null
76dc84a5a805a2c8b91e9ccc00b8dbf8f4a99bf0d56ab25832f6e6addd4f7f57
"{\"include_stop_str_in_output\":false,\"logprobs\":128,\"max_model_len\":12288,\"max_prompt_tokens\(...TRUNCATED)
"{\"platform\":\"Linux-6.8.0-1046-nvidia-x86_64-with-glibc2.35\",\"pyarrow\":\"25.0.0\",\"python\":\(...TRUNCATED)
hf_bf16_sdpa_full_forward
bde9e800223cdd62228ce39e0305398f6ada05b98adaf438b0b3d3d3c3015561
bfloat16
sdpa
30
2026-07-31T06:26:16.720095+00:00
gemma4-hf-bf16-sdpa-topk-overlay-v1
1c409d348a85f8d66941e7cb0700279af127147d82c39e6079aab929fcfca17a
efe76f5a53225e97081a750495ceb8ebe26b1a72a3ea8e0b74ea079a87828c0a
17ba236cfe2fdc25ed6f2d5e145534abfdd890a5d42396e8b801012301a52f64
a9c4d270f0246252b6e4d3478f99af61b7f41e78e552d76a43025a0fca55bd78
a2c1aacb4166472a6da7c0af223270b0feb289c1d612b08ea5c090184687d75c
7ef64928a49efd0da944612b66956dd24bdd8572b0226d2959b633b9f9c1c59d
e2b_base_to_e4b
train
deepscaler_4of4strict_rl_train
407211b033fb4013e7199012f441301e4012caf9d493a1321587bdae3c894918
row:00000001:sha256:efc45d868ae38cb3ce44a034098717d857d2e5d9b4acdfa4b8947aabf8bd3387
efc45d868ae38cb3ce44a034098717d857d2e5d9b4acdfa4b8947aabf8bd3387
1
0
How many perfect squares are between 50 and 250? Please output the final answer within \boxed{}.
8
0
false
4
"We can check whether a number is a perfect square by finding all the square roots of its perfect sq(...TRUNCATED)
"We can check whether a number is a perfect square by finding all the square roots of its perfect sq(...TRUNCATED)
[2,236820,3041,236779,1340,236779,887,236813,2364,107,11823,506,8610,529,506,5619,874,1420,1564,4784(...TRUNCATED)
[1882,740,2426,3363,496,1548,563,496,4011,6281,684,8159,784,506,6281,14546,529,1061,4011,6281,5872,2(...TRUNCATED)
[2,236820,3041,236779,1340,236779,887,236813,2364,107,11823,506,8610,529,506,5619,874,1420,1564,4784(...TRUNCATED)
[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0(...TRUNCATED)
[[3810,818,1882,236795,34531,54133,236810,10081,2021,7280,3292,6481,142612,236776,43812,3243,4663,76(...TRUNCATED)
[[-1.41796875,-1.54296875,-2.04296875,-2.66796875,-3.04296875,-3.66796875,-3.79296875,-4.29296875,-4(...TRUNCATED)
[-2.04296875,-2.1484375,-3.013671875,-3.591796875,-0.63232421875,-0.2386474609375,-0.299560546875,-0(...TRUNCATED)
1..128
1,548
123
0
5
vllm
/tmp/hf_cache/models--google--gemma-4-E2B/snapshots/d29ff6b45f081a49ee2733a859c9c9c2d95d1a6f
null
76dc84a5a805a2c8b91e9ccc00b8dbf8f4a99bf0d56ab25832f6e6addd4f7f57
"{\"include_stop_str_in_output\":false,\"logprobs\":128,\"max_model_len\":12288,\"max_prompt_tokens\(...TRUNCATED)
"{\"platform\":\"Linux-6.8.0-1046-nvidia-x86_64-with-glibc2.35\",\"pyarrow\":\"25.0.0\",\"python\":\(...TRUNCATED)
hf_bf16_sdpa_full_forward
bde9e800223cdd62228ce39e0305398f6ada05b98adaf438b0b3d3d3c3015561
bfloat16
sdpa
30
2026-07-31T06:26:16.720095+00:00
gemma4-hf-bf16-sdpa-topk-overlay-v1
1c409d348a85f8d66941e7cb0700279af127147d82c39e6079aab929fcfca17a
efe76f5a53225e97081a750495ceb8ebe26b1a72a3ea8e0b74ea079a87828c0a
17ba236cfe2fdc25ed6f2d5e145534abfdd890a5d42396e8b801012301a52f64
a9c4d270f0246252b6e4d3478f99af61b7f41e78e552d76a43025a0fca55bd78
a2c1aacb4166472a6da7c0af223270b0feb289c1d612b08ea5c090184687d75c
59baceaf3b106672baa1d1bc85ee76634f4a992c8b12b19b6dc412d60dd9ae55
e2b_base_to_e4b
train
deepscaler_4of4strict_rl_train
407211b033fb4013e7199012f441301e4012caf9d493a1321587bdae3c894918
row:00000001:sha256:efc45d868ae38cb3ce44a034098717d857d2e5d9b4acdfa4b8947aabf8bd3387
efc45d868ae38cb3ce44a034098717d857d2e5d9b4acdfa4b8947aabf8bd3387
1
1
How many perfect squares are between 50 and 250? Please output the final answer within \boxed{}.
8
0
false
4 (\sqrt{5} - \sqrt{2})
"We can factor 50 as $2^2 \\cdot 5 \\Rightarrow 50 - \\sqrt{50} = 4 \\sqrt{2}$, and factor 250 as $2(...TRUNCATED)
"We can factor 50 as $2^2 \\cdot 5 \\Rightarrow 50 - \\sqrt{50} = 4 \\sqrt{2}$, and factor 250 as $2(...TRUNCATED)
[2,236820,3041,236779,1340,236779,887,236813,2364,107,11823,506,8610,529,506,5619,874,1420,1564,4784(...TRUNCATED)
[1882,740,5415,236743,236810,236771,618,609,236778,236884,236778,621,4172,236743,236810,621,13676,23(...TRUNCATED)
[2,236820,3041,236779,1340,236779,887,236813,2364,107,11823,506,8610,529,506,5619,874,1420,1564,4784(...TRUNCATED)
[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0(...TRUNCATED)
[[3810,818,1882,236795,34531,54133,236810,10081,2021,7280,3292,6481,142612,236776,43812,3243,4663,76(...TRUNCATED)
[[-1.41796875,-1.54296875,-2.04296875,-2.66796875,-3.04296875,-3.66796875,-3.79296875,-4.29296875,-4(...TRUNCATED)
[-2.04296875,-2.1484375,-5.13671875,-0.72021484375,-0.60546875,-0.006683349609375,-0.349853515625,-0(...TRUNCATED)
1..128
1,548
142
0
6
vllm
/tmp/hf_cache/models--google--gemma-4-E2B/snapshots/d29ff6b45f081a49ee2733a859c9c9c2d95d1a6f
null
76dc84a5a805a2c8b91e9ccc00b8dbf8f4a99bf0d56ab25832f6e6addd4f7f57
"{\"include_stop_str_in_output\":false,\"logprobs\":128,\"max_model_len\":12288,\"max_prompt_tokens\(...TRUNCATED)
"{\"platform\":\"Linux-6.8.0-1046-nvidia-x86_64-with-glibc2.35\",\"pyarrow\":\"25.0.0\",\"python\":\(...TRUNCATED)
hf_bf16_sdpa_full_forward
bde9e800223cdd62228ce39e0305398f6ada05b98adaf438b0b3d3d3c3015561
bfloat16
sdpa
30
2026-07-31T06:26:16.720095+00:00
gemma4-hf-bf16-sdpa-topk-overlay-v1
1c409d348a85f8d66941e7cb0700279af127147d82c39e6079aab929fcfca17a
efe76f5a53225e97081a750495ceb8ebe26b1a72a3ea8e0b74ea079a87828c0a
17ba236cfe2fdc25ed6f2d5e145534abfdd890a5d42396e8b801012301a52f64
a9c4d270f0246252b6e4d3478f99af61b7f41e78e552d76a43025a0fca55bd78
a2c1aacb4166472a6da7c0af223270b0feb289c1d612b08ea5c090184687d75c
32f2e30f038ea6fbe34adb1806b94b37dff05551d12bc32939131ce9dd1ea397
e2b_base_to_e4b
train
deepscaler_4of4strict_rl_train
407211b033fb4013e7199012f441301e4012caf9d493a1321587bdae3c894918
row:00000001:sha256:efc45d868ae38cb3ce44a034098717d857d2e5d9b4acdfa4b8947aabf8bd3387
efc45d868ae38cb3ce44a034098717d857d2e5d9b4acdfa4b8947aabf8bd3387
1
2
How many perfect squares are between 50 and 250? Please output the final answer within \boxed{}.
8
0
false
57
"We are going to find $\\sqrt{250}, \\sqrt{249}$ to learn the pattern. $\\sqrt{250} < 5\\sqrt{50} < (...TRUNCATED)
"We are going to find $\\sqrt{250}, \\sqrt{249}$ to learn the pattern. $\\sqrt{250} < 5\\sqrt{50} < (...TRUNCATED)
[2,236820,3041,236779,1340,236779,887,236813,2364,107,11823,506,8610,529,506,5619,874,1420,1564,4784(...TRUNCATED)
[1882,659,1771,531,1586,874,4784,236782,236778,236810,236771,1263,621,4784,236782,236778,236812,2368(...TRUNCATED)
[2,236820,3041,236779,1340,236779,887,236813,2364,107,11823,506,8610,529,506,5619,874,1420,1564,4784(...TRUNCATED)
[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0(...TRUNCATED)
[[3810,818,1882,236795,34531,54133,236810,10081,2021,7280,3292,6481,142612,236776,43812,3243,4663,76(...TRUNCATED)
[[-1.41796875,-1.54296875,-2.04296875,-2.66796875,-3.04296875,-3.66796875,-3.79296875,-4.29296875,-4(...TRUNCATED)
[-2.04296875,-2.7734375,-4.6171875,-0.265380859375,-1.794921875,-4.3203125,-0.32861328125,-0.0828247(...TRUNCATED)
1..128
1,548
121
0
7
vllm
/tmp/hf_cache/models--google--gemma-4-E2B/snapshots/d29ff6b45f081a49ee2733a859c9c9c2d95d1a6f
null
76dc84a5a805a2c8b91e9ccc00b8dbf8f4a99bf0d56ab25832f6e6addd4f7f57
"{\"include_stop_str_in_output\":false,\"logprobs\":128,\"max_model_len\":12288,\"max_prompt_tokens\(...TRUNCATED)
"{\"platform\":\"Linux-6.8.0-1046-nvidia-x86_64-with-glibc2.35\",\"pyarrow\":\"25.0.0\",\"python\":\(...TRUNCATED)
hf_bf16_sdpa_full_forward
bde9e800223cdd62228ce39e0305398f6ada05b98adaf438b0b3d3d3c3015561
bfloat16
sdpa
30
2026-07-31T06:26:16.720095+00:00
gemma4-hf-bf16-sdpa-topk-overlay-v1
1c409d348a85f8d66941e7cb0700279af127147d82c39e6079aab929fcfca17a
efe76f5a53225e97081a750495ceb8ebe26b1a72a3ea8e0b74ea079a87828c0a
17ba236cfe2fdc25ed6f2d5e145534abfdd890a5d42396e8b801012301a52f64
a9c4d270f0246252b6e4d3478f99af61b7f41e78e552d76a43025a0fca55bd78
a2c1aacb4166472a6da7c0af223270b0feb289c1d612b08ea5c090184687d75c
cc0b1f39e64ec9ff9286ed098856a729be6d3d2eb26adf931319fba6e70a7333
e2b_base_to_e4b
train
deepscaler_4of4strict_rl_train
407211b033fb4013e7199012f441301e4012caf9d493a1321587bdae3c894918
row:00000001:sha256:efc45d868ae38cb3ce44a034098717d857d2e5d9b4acdfa4b8947aabf8bd3387
efc45d868ae38cb3ce44a034098717d857d2e5d9b4acdfa4b8947aabf8bd3387
1
3
How many perfect squares are between 50 and 250? Please output the final answer within \boxed{}.
8
0
false
"You can list all whole perfect squares between 50 and 250.<style> Please annotate each perfect squa(...TRUNCATED)
"You can list all whole perfect squares between 50 and 250.<style> Please annotate each perfect squa(...TRUNCATED)
[2,236820,3041,236779,1340,236779,887,236813,2364,107,11823,506,8610,529,506,5619,874,1420,1564,4784(...TRUNCATED)
[3048,740,1694,784,3697,4011,23441,1534,236743,236810,236771,532,236743,236778,236810,236771,21603,3(...TRUNCATED)
[2,236820,3041,236779,1340,236779,887,236813,2364,107,11823,506,8610,529,506,5619,874,1420,1564,4784(...TRUNCATED)
[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0(...TRUNCATED)
[[3810,818,1882,236795,34531,54133,236810,10081,2021,7280,3292,6481,142612,236776,43812,3243,4663,76(...TRUNCATED)
[[-1.41796875,-1.54296875,-2.04296875,-2.66796875,-3.04296875,-3.66796875,-3.79296875,-4.29296875,-4(...TRUNCATED)
[-6.66796875,-0.7646484375,-2.658203125,-1.5615234375,-7.12109375,-1.244140625,-0.03167724609375,-0.(...TRUNCATED)
1..128
1,548
152
0
8
vllm
/tmp/hf_cache/models--google--gemma-4-E2B/snapshots/d29ff6b45f081a49ee2733a859c9c9c2d95d1a6f
null
76dc84a5a805a2c8b91e9ccc00b8dbf8f4a99bf0d56ab25832f6e6addd4f7f57
"{\"include_stop_str_in_output\":false,\"logprobs\":128,\"max_model_len\":12288,\"max_prompt_tokens\(...TRUNCATED)
"{\"platform\":\"Linux-6.8.0-1046-nvidia-x86_64-with-glibc2.35\",\"pyarrow\":\"25.0.0\",\"python\":\(...TRUNCATED)
hf_bf16_sdpa_full_forward
bde9e800223cdd62228ce39e0305398f6ada05b98adaf438b0b3d3d3c3015561
bfloat16
sdpa
30
2026-07-31T06:26:16.720095+00:00
gemma4-hf-bf16-sdpa-topk-overlay-v1
1c409d348a85f8d66941e7cb0700279af127147d82c39e6079aab929fcfca17a
efe76f5a53225e97081a750495ceb8ebe26b1a72a3ea8e0b74ea079a87828c0a
17ba236cfe2fdc25ed6f2d5e145534abfdd890a5d42396e8b801012301a52f64
a9c4d270f0246252b6e4d3478f99af61b7f41e78e552d76a43025a0fca55bd78
a2c1aacb4166472a6da7c0af223270b0feb289c1d612b08ea5c090184687d75c
497d85382ef58d72af944fd42294f54c5fda5c723b9de98f01116bc8c8ae533a
e2b_base_to_e4b
train
deepscaler_4of4strict_rl_train
407211b033fb4013e7199012f441301e4012caf9d493a1321587bdae3c894918
row:00000001:sha256:efc45d868ae38cb3ce44a034098717d857d2e5d9b4acdfa4b8947aabf8bd3387
efc45d868ae38cb3ce44a034098717d857d2e5d9b4acdfa4b8947aabf8bd3387
1
4
How many perfect squares are between 50 and 250? Please output the final answer within \boxed{}.
8
0
false
"50 is $25^2$, 56 is $25^2 + 1$, 67 is $25^2 + 9$, 78 is $25^2 + 16$, 89 is $25^2 + 26$, 90 is $25^2(...TRUNCATED)
"50 is $25^2$, 56 is $25^2 + 1$, 67 is $25^2 + 9$, 78 is $25^2 + 16$, 89 is $25^2 + 26$, 90 is $25^2(...TRUNCATED)
[2,236820,3041,236779,1340,236779,887,236813,2364,107,11823,506,8610,529,506,5619,874,1420,1564,4784(...TRUNCATED)
[236810,236771,563,609,236778,236810,236884,236778,2500,236743,236810,236825,563,609,236778,236810,2(...TRUNCATED)
[2,236820,3041,236779,1340,236779,887,236813,2364,107,11823,506,8610,529,506,5619,874,1420,1564,4784(...TRUNCATED)
[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0(...TRUNCATED)
[[3810,818,1882,236795,34531,54133,236810,10081,2021,7280,3292,6481,142612,236776,43812,3243,4663,76(...TRUNCATED)
[[-1.41796875,-1.54296875,-2.04296875,-2.66796875,-3.04296875,-3.66796875,-3.79296875,-4.29296875,-4(...TRUNCATED)
[-3.79296875,-0.15625,-1.14453125,-1.974609375,-0.62255859375,-0.6455078125,-0.19775390625,-0.001296(...TRUNCATED)
1..128
1,548
143
0
9
vllm
/tmp/hf_cache/models--google--gemma-4-E2B/snapshots/d29ff6b45f081a49ee2733a859c9c9c2d95d1a6f
null
76dc84a5a805a2c8b91e9ccc00b8dbf8f4a99bf0d56ab25832f6e6addd4f7f57
"{\"include_stop_str_in_output\":false,\"logprobs\":128,\"max_model_len\":12288,\"max_prompt_tokens\(...TRUNCATED)
"{\"platform\":\"Linux-6.8.0-1046-nvidia-x86_64-with-glibc2.35\",\"pyarrow\":\"25.0.0\",\"python\":\(...TRUNCATED)
hf_bf16_sdpa_full_forward
bde9e800223cdd62228ce39e0305398f6ada05b98adaf438b0b3d3d3c3015561
bfloat16
sdpa
30
2026-07-31T06:26:16.720095+00:00
End of preview. Expand in Data Studio

Gemma 4 E2B base top-k-128 HF training overlay

This is the immutable training-engine overlay used to distill traces from Gemma 4 E2B base into Gemma 4 E4B. It preserves the prompts, responses, and exact response token IDs from JWei05/gemma4-e2b-base-topk128-traces, but replaces the source vLLM top-k targets with targets recomputed by the Hugging Face training engine.

This repository is a reproducibility artifact for the corresponding distillation run. It is not a new independently collected corpus.

Contents

Split Rows Response tokens Shards
train 48,615 8,087,407 1,216
validation 128 18,437 16
total 48,743 8,105,844 1,232

The validation split contains one trace from each of 128 deterministic, unique, train-disjoint validation prompts selected with seed 42. The Parquet shards use int32 token IDs and float16 stored log probabilities.

Target computation

  • Teacher: google/gemma-4-E2B snapshot d29ff6b45f081a49ee2733a859c9c9c2d95d1a6f
  • Engine: unsharded Hugging Face Transformers forward
  • Compute dtype: BF16
  • Attention implementation: SDPA
  • Vocabulary normalization: FP32 full-vocabulary logsumexp after the BF16 language-model head and final-logit softcap
  • Top-k width: 128 over a 262,144-token vocabulary
  • Alignment: response token at input position i is scored by logits at i - 1
  • Maximum sequence length: 12,288 tokens

The original response token IDs are never re-tokenized. rescore_config.json, parity_receipt.json, and dataset_index.json record the complete semantic configuration and integrity identities.

Integrity identities

  • Dataset index SHA-256: 124a1b904b60963fb2b1d422107bec593a8ef1053cce34fff40bfb6314d1a16e
  • Source dataset index SHA-256: efe76f5a53225e97081a750495ceb8ebe26b1a72a3ea8e0b74ea079a87828c0a
  • Rescore configuration SHA-256: 1c409d348a85f8d66941e7cb0700279af127147d82c39e6079aab929fcfca17a
  • Teacher weight-content SHA-256: 76dc84a5a805a2c8b91e9ccc00b8dbf8f4a99bf0d56ab25832f6e6addd4f7f57
  • Teacher model-identity SHA-256: bde9e800223cdd62228ce39e0305398f6ada05b98adaf438b0b3d3d3c3015561

Intended use

Use this overlay when reproducing the E2B-base-to-E4B top-k distillation experiment with the repository's gemma4_topk_distill_fsdp2.sh launcher. Consumers should validate dataset_index.json before training and preserve the registered shard order.

The generated traces and probabilities remain subject to the terms governing the source dataset and Gemma models. Review those terms before redistribution or downstream use.

Downloads last month
19

Models trained or fine-tuned on JWei05/gemma4-e2b-base-topk128-hf-overlay-v128-seed42