input_ids list |
|---|
[
248045,
846,
198,
7734,
264,
220,
17,
15,
15,
15,
36093,
17832,
3255,
883,
264,
1809,
314,
4097,
194978,
1663,
11,
1754,
440,
836,
4752,
17020,
11,
852,
67080,
383,
264,
57288,
757,
11285,
310,
3462,
836,
93954,
13245,
494,
279,
60356... |
[
5273,
220,
16,
25,
31304,
321,
12387,
2193,
13,
34677,
3200,
220,
16,
24,
337,
14,
21453,
220,
17,
15,
337,
1478,
25089,
11,
11308,
77935,
11,
42152,
11,
19850,
1619,
11,
7277,
24479,
11,
35343,
7431,
11,
1675,
45585,
15203,
11567,
... |
[
248045,
8678,
198,
24342,
286,
4879,
369,
716,
310,
830,
11553,
13,
5044,
1683,
15060,
1472,
279,
3274,
11,
9307,
1328,
30800,
11,
2814,
47675,
25605,
11,
321,
60445,
55404,
11,
27224,
11,
321,
30246,
303,
279,
1534,
4087,
13,
271,
2,... |
[
279,
1412,
1500,
369,
494,
533,
7116,
276,
25,
1510,
66,
1402,
2226,
66,
17,
3359,
2226,
66,
3359,
2226,
66,
63073,
834,
25,
29116,
25,
18,
20,
25,
15135,
1412,
25,
2706,
314,
819,
7210,
314,
913,
359,
3504,
514,
6,
63,
321,
151... |
[
21,
19,
14813,
25903,
13,
13875,
4011,
16000,
413,
6206,
369,
2166,
3349,
321,
28782,
888,
1510,
3946,
3456,
40061,
63,
318,
3946,
314,
220,
18,
17,
14813,
4105,
8,
41915,
13,
271,
4987,
803,
1245,
21541,
12846,
2612,
6206,
11,
524,
... |
[
7156,
33875,
369,
1870,
913,
13,
733,
21123,
273,
1510,
8090,
33875,
283,
804,
63,
466,
958,
7156,
1870,
13,
14235,
364,
1542,
31363,
13,
198,
12,
26567,
421,
13514,
1510,
36047,
63,
440,
12062,
3817,
321,
1510,
1643,
63,
303,
1510,
... |
[
2099,
15105,
13,
198,
24,
13,
71680,
28133,
2276,
22414,
1590,
1248,
3644,
13,
11760,
478,
328,
12516,
46018,
55615,
1,
7612,
3037,
13,
42123,
430,
83999,
318,
2359,
1528,
681,
220,
21,
13067,
23985,
3835,
21,
12,
23,
17820,
11,
1754,... |
[
30,
2717,
2098,
13,
733,
45309,
11,
281,
2784,
8142,
12435,
16,
27736,
11,
41548,
8,
628,
599,
830,
16,
23478,
13,
733,
716,
42727,
11,
413,
781,
1910,
328,
16,
15,
17,
19,
11,
220,
17,
15,
16,
21,
11,
830,
62,
16,
27736,
11,
... |
[
1786,
10810,
43824,
440,
264,
2609,
314,
8851,
1278,
87,
459,
310,
7005,
777,
11044,
7620,
2519,
16916,
34128,
19,
47348,
3762,
375,
321,
43187,
13,
198,
18998,
2519,
10810,
369,
9722,
279,
3460,
321,
6676,
314,
279,
10810,
1814,
13,
... |
[
310,
1910,
264,
855,
579,
3299,
369,
279,
1132,
15787,
421,
3070,
914,
18725,
1149,
34503,
3659,
780,
23213,
30,
1216,
5533,
8619,
310,
1791,
73655,
25,
328,
9020,
416,
303,
1146,
3187,
1288,
1179,
8346,
14677,
30,
1921,
413,
551,
579... |
[
12848,
302,
754,
11,
16976,
11,
29403,
2528,
198,
18,
13,
31169,
21930,
7432,
1892,
152688,
118,
16,
17,
13,
20,
15,
15,
14,
9011,
1892,
220,
24,
21,
295,
28495,
11,
843,
7087,
11,
4204,
694,
1482,
11,
303,
46150,
17378,
11,
843,
... |
[
30,
32645,
3311,
11629,
220,
19,
87,
19,
30,
14235,
3799,
30,
1368,
25545,
303,
26684,
18901,
11,
1144,
3799,
9727,
2094,
3311,
11629,
13,
28763,
25,
220,
16,
25545,
14,
16828,
220,
19,
13,
17,
20,
303,
830,
220,
19,
303,
13,
1038... |
[
1987,
364,
423,
28,
17,
15,
11,
2702,
2438,
283,
17,
15,
23197,
307,
16,
37063,
18,
28,
17,
13,
8439,
307,
18,
85514,
314,
220,
18,
2438,
82,
13,
7054,
423,
2526,
11,
4507,
874,
220,
18,
30559,
22983,
30,
16018,
381,
13,
1368,
... |
[
369,
874,
1144,
310,
1925,
264,
491,
1906,
13,
198,
510,
33639,
29,
248046,
198,
248045,
74455,
198,
248068,
271,
248069,
271,
12675,
0,
2500,
628,
353,
1438,
488,
3242,
30,
248046,
198,
248045,
846,
198,
12675,
1017,
0,
353,
1144,
31... |
[
11782,
198,
12,
9321,
25,
8239,
33877,
709,
303,
20068,
14,
6281,
25,
1870,
1411,
24446,
1833,
1411,
1777,
1411,
3305,
13,
35032,
57018,
25,
24446,
1833,
3945,
13,
198,
12,
18284,
10814,
25,
198,
220,
220,
16,
13,
13086,
286,
1503,
... |
[
2577,
2094,
220,
23,
15,
15,
13,
6558,
579,
1698,
4105,
539,
13901,
1608,
11048,
13,
271,
3723,
25,
328,
760,
13237,
73473,
1,
198,
10561,
1698,
6970,
13,
271,
48401,
25,
198,
9726,
13237,
73473,
25,
8169,
279,
6886,
37750,
3114,
27... |
[
248045,
846,
198,
27289,
264,
20795,
3255,
883,
264,
1809,
314,
4627,
852,
71616,
5028,
264,
60755,
1576,
13,
248046,
198,
248045,
74455,
198,
248068,
271,
248069,
271,
12162,
5028,
264,
854,
11,
1017,
998,
2943,
4627,
6725,
17926,
11,
... |
[
321,
5090,
975,
13565,
321,
9919,
3882,
888,
21707,
13,
198,
7732,
10682,
9257,
303,
279,
26646,
62923,
2094,
279,
26312,
314,
32274,
303,
279,
9673,
11,
353,
18397,
279,
26312,
0,
9322,
369,
2414,
353,
4021,
11,
748,
567,
2224,
2019,... |
[
248045,
74455,
198,
248068,
198,
50525,
25,
411,
2313,
11222,
12654,
2518,
220,
18,
13,
23,
27718,
40007,
69486,
283,
328,
9628,
18,
13,
23,
1,
7561,
321,
821,
3181,
369,
1048,
19152,
440,
1510,
12305,
9159,
283,
328,
18,
13,
23,
38... |
[
1087,
1395,
4175,
785,
10288,
1284,
928,
53963,
12,
1341,
18140,
44095,
6604,
738,
29540,
82,
10288,
12,
55383,
8404,
1395,
350,
908,
16575,
903,
2074,
3309,
1723,
4073,
20599,
9485,
12,
1449,
2777,
49097,
198,
1851,
4501,
25,
16932,
22... |
[
12,
19,
13,
16,
13,
15,
63,
30,
1921,
413,
781,
4722,
220,
19,
13,
16,
13,
17,
9052,
310,
220,
19,
13,
16,
13,
16,
4722,
30,
14235,
9496,
13,
426,
3310,
8240,
25,
4581,
11,
10582,
6942,
318,
87,
2307,
13,
15,
681,
3555,
2772... |
CatTrellis Seed
The seed data of CatTrellis Stage 0 (cattrellis.stage0). Stage 0
builds a model's Stage 1 calibration rows ([2048, 2048]) and Stage 2 training rows ([4096, 4096]) from this
dataset:
- Generated pools (chat, math/STEM and cyber reasoning, agent turns, MiMo RL tasks): the model being quantized
answers the seed prompts in
calib-v4/prompts/in vLLM, so these rows are its own completions. - Copied pools (C4, UltraChat, on-policy SWE-agent traces, long agentic traces): taken from the two published matrices below as they are, re-tokenized for a model with a vocabulary other than Qwen3.8-27B's.
The matrices also fix the layout: rows per pool, and which pools are generated and which copied. Their generated pools' rows (Qwen3.8-27B completions) are rebuilt from the model's own completions and not used. All token ids are those of the Qwen/Qwen3.8-27B tokenizer.
Subsets
calib-v4: the seed prompts and the calibration matrix
| file | content |
|---|---|
prompts/chat.jsonl |
1,200 chat prompts (half in thinking mode) |
prompts/rl.jsonl |
1,100 verifiable tasks: 700 math, 400 STEM knowledge |
prompts/cyber.jsonl |
600 security reasoning prompts |
prompts/uagent.jsonl |
700 real agent-trajectory prefixes (400 tool use, 150 code, 150 search), cut before an assistant turn that the model writes |
prompts/mimo.jsonl |
1,200 RL environment tasks (376 code, 275 general, 238 webdev, 177 cyber, 134 music), with their tools |
calib-4096x2048.parquet |
[4096, 2048] token ids (input_ids: int32 fixed-length list); the published calibration matrix |
calib-4096x2048.json |
its manifest: rows per pool, row_sources (the pool of each row), per-pool provenance, sha256 (over the int64 bytes) |
A prompt record is {uid, domain, mode, split, messages, tools, enable_thinking}: chat-template messages (and
tools schemas) rendered with the model's own chat template, enable_thinking=false on the non-thinking chat
prompts. The five files have different message and tool schemas, so they are plain JSONL files rather than a
datasets config; read them with json:
import json
from huggingface_hub import hf_hub_download
path = hf_hub_download("casperhansen/CatTrellis-Seed", "calib-v4/prompts/rl.jsonl", repo_type="dataset")
prompts = [json.loads(line) for line in open(path)]
Calibration pools (rows of 2,048 tokens):
| pool | rows | kind | content |
|---|---|---|---|
c4 |
384 | copied | C4 web text, random 2,048-token spans |
ultrachat |
256 | copied | real UltraChat conversations (non-thinking), packed |
chat4 |
384 | generated | prompts/chat.jsonl |
rl4 |
640 | generated | prompts/rl.jsonl |
cyber4 |
320 | generated | prompts/cyber.jsonl |
agent4 |
384 | generated | prompts/uagent.jsonl (the real prefix, then the model's turn) |
mimo4 |
704 | generated | prompts/mimo.jsonl |
swe4 |
640 | copied | Open-SWE-Traces trajectories written by Qwen3.8-27B, one window per trace |
longchop4 |
384 | copied | 12 long traces (65,536 tokens each: 6 SWE, 2 UltraData code agent, 2 long-document QA, 2 long agent continuations) chopped into consecutive rows |
stage3-data: the Stage 2 matrix and the held-out sets
| file | content |
|---|---|
train/mix-d6-4096x4096.parquet |
[4096, 4096] token ids (input_ids: int32 fixed-length list); the published Stage 2 matrix ("d6") |
train/mix-d6-4096x4096.json |
its manifest: parts (v4_at_4096: the calibration pools re-cut at 4,096; new_batch: a new batch of the same pool types), rows, row_sources, per-pool provenance, sha256 |
heldout/heldout-c4-64x2048.npy |
64 x 2,048 C4 tokens, read by tools/kl_eval.py |
heldout/heldout-reason-16x2048.npy |
16 x 2,048 tokens of Qwen3.8-27B thinking traces on held-out prompts |
heldout/heldout-long-16x8192.npy (+ .json) |
16 x 8,192 tokens of long thinking rollouts and on-policy SWE traces; the .json lists each row's source |
eval/roll_fp_math.jsonl, eval/roll_fp_agent.jsonl |
64 math and 32 agent rollouts of BF16 Qwen3.8-27B (prompt_ids, completion_ids) |
eval/agent_teacher_64k.jsonl |
64 BF16 Qwen3.8-27B episodes on the 32 held-out agent tasks at 64k context, read by tools/mimo.py |
Stage 2's copied new-batch pools are c4n (C4 train shard 00004), ultrachatn (UltraChat train_sft shard 0
conversations not used elsewhere), swen (Open-SWE-Traces scale-swe shard 00002) and longn (12 new long traces);
its generated new-batch pools (chatn, rln, cybern, agentn, mimon) take the prompts after those the
calibration pools used, in the same files.
Every held-out and evaluation set is disjoint from the training prompts and rows (different splits, records, shards or tasks).
Load a matrix with:
import datasets, numpy as np, torch
ds = datasets.load_dataset("casperhansen/CatTrellis-Seed", "calib-v4", split="train") # or "stage3-data"
ids = torch.from_numpy(ds.with_format("numpy")["input_ids"][:].astype(np.int64)) # [4096, 2048] int64
Where the seed prompts come from
The upstream links are pinned to a revision. The prompt files were built before these pins were recorded, so each pin is the revision the prompts were checked against: every prompt's text matches its source record there.
| prompt file | source | license | how the prompts were taken |
|---|---|---|---|
chat.jsonl |
HuggingFaceH4/ultrachat_200k, train_sft shard 0 |
MIT | the first user turn of a seeded random draw of conversations; uid is uc-<prompt_id>; half are answered in thinking mode |
rl.jsonl (math) |
openbmb/UltraData-RL-2609, data/Math/Math_part-1-of-4.jsonl |
Apache-2.0 | the record's query as one user turn; uid is rl-math-<uuid> |
rl.jsonl (knowledge) |
openbmb/UltraData-RL-2609, data/Knowledge/Knowledge_part-1-of-2.jsonl |
Apache-2.0 | as above; uid is rl-knowledge-<uuid> |
cyber.jsonl |
echel0nn1881/kimi-cyber-reasoning, train.jsonl records 0-899 (900 onward are held out) |
WTFPL | the system and user turns before the first assistant turn; uid is cyber-<record index> |
uagent.jsonl |
openbmb/UltraData-SFT-Agent-2609: Tool_Use part 1, Code_Agent part 1, Search_Agent part 1 |
Apache-2.0 | a trajectory (teacher turns and real tool outputs) cut before a random assistant turn, 128-8,192 tokens of prefix; the last 12 records of each shard are held out |
mimo.jsonl |
XiaomiMiMo/MiMo-V2.6-RL-oss at 639865fd33: code.parquet, general/train.parquet, webdev.parquet, cyber.parquet, music.parquet |
Apache-2.0 | the task's prompt messages; uid is <domain>-<row index>. Except for music and tasks that bring their own, a CatTrellis system prompt is prepended: an agent prompt with file/shell tool schemas (mode: tools) or a solve-it-in-one-reply prompt (mode: solve). The code-<i> row indices refer to code.parquet at this revision: the next upstream commit (6b9046cc42) dropped 96 tasks and shifted the rows |
Where the copied rows come from
| pool | source | license |
|---|---|---|
c4, c4n, heldout-c4 |
allenai/c4 en: train shard 00000 (c4), train shard 00004 (c4n) |
ODC-BY 1.0 |
ultrachat, ultrachatn |
HuggingFaceH4/ultrachat_200k train_sft shard 0 |
MIT |
swe4, swen, SWE traces of longchop4 / longn / heldout-long |
nvidia/Open-SWE-Traces minisweagent/qwen38_27b: scale-swe shard 00001 (swe4, longchop4), scale-swe shard 00002 (swen, longn), swe-rebench-v2 (held out) |
CC-BY-4.0 |
code agent traces of longchop4 |
openbmb/UltraData-SFT-Agent-2609 Code_Agent part 1 (teacher trajectories) |
Apache-2.0 |
long agent continuations of longchop4 / longn |
openbmb/UltraData-SFT-Agent-2609 General_Agent and Search_Agent prefixes of 16-30k tokens, continued by Qwen3.8-27B |
Apache-2.0 |
long-document QA of longchop4 / longn |
openbmb/UltraData-RL-2609 Long_Context part 1 questions, answered by Qwen3.8-27B |
Apache-2.0 |
| model-written text in every pool and set | Qwen/Qwen3.8-27B generations (BF16, vLLM, its generation_config sampling) |
per the model's license |
See LICENSE.md. C4 (ODC-BY) and Open-SWE-Traces (CC-BY-4.0) require attribution; the SWE trajectories
contain code from public repositories under their own licenses (the license field of the source dataset).
History
The files are byte-for-byte copies of two earlier private sets: calib-v4/ from casperhansen/dwarftree-calib-v4
at 79c8948f19751e621ed6b6ae56b0d500e6241790 (its prompt files and calibration matrix) and stage3-data/ from
casperhansen/dwarftree-stage3-data at cf41278fa7dbc3aa01a4385446f77ffbf79ab370 (train/mix-d6-4096x4096.*,
heldout/, eval/). CatTrellis was called DwarfTree then.
- Downloads last month
- -