Hugging Face's logo

Datasets:
IO-AIR
/
filterme-sft

Tasks:
Question Answering
Text Generation
Modalities:
Tabular
Text
Formats:
json
Size:
100K - 1M
Tags:
sft
instruction-tuning
agentic
reverse-engineering
cad
svg
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
filterme-sft / README.md
IO-AIR's picture
IO-AIR
bigbuild v2.0 finals: README.md
d575c97 verified
preview code
|
Raw History Blame Contribute Delete
5.97 kB
metadata 
license: cc-by-4.0
task_categories:
  - question-answering
  - text-generation
  - text2text-generation
tags:
  - sft
  - instruction-tuning
  - agentic
  - reverse-engineering
  - cad
  - svg
  - long-context
  - synthetic
source_datasets: sft-forge
pretty_name: sft-forge ensemble v2.0

sft-forge-ensemble v2.0

One pareto-pruned, token-rebalanced SFT mixture built by sft-forge. 321,777 rows · 2,123,529,634 real gpt2 tokens · fine-grained tags on every row.

Token mix by domain (real gpt2, kept rows)

domain gpt2 tokens share
reasoning_math 551,186,692 26.0%
reasoning_sci 370,397,475 17.4%
agentic_swe 304,253,237 14.3%
code_agent 233,702,388 11.0%
agentic_tool 219,243,982 10.3%
general_sft 213,467,470 10.1%
reverse_engineering 147,388,757 6.9%
vision_svg 67,429,930 3.2%
cad_3d 9,174,147 0.4%
vision_plot 4,315,416 0.2%
blender_3d 1,220,418 0.1%
godot_dev 929,412 0.0%
vision_image 820,310 0.0%

Sources

source rows gpt2 tokens
openmath_cot 66787 535,300,110
opensci 65779 389,184,121
openmath_tir 136459 1,190,615,653
openr1_math 38000 224,337,795
nemotron_sci 35000 52,208,521
ultrachat 62485 75,454,571
svg_code 14317 7,279,031
mcp_atlas 462 7,583,866
pyromind_long 1000 58,991,782
pyromind_short 1932 14,639,677
toolmind 13351 171,755,540
ultradata_tooluse 2095 13,520,358
nemotron_swe 4702 271,084,502
swesmith 5174 164,528,053
coderforge 2184 199,432,411
nemotron_cuda 2276 23,010,820
ultradata_code 4200 276,443,679
agenticsft 500 8,115,596
demyagent 500 1,179,050
ultradata_general 2692 69,243,262
ultradata_search 2576 76,121,807
ollvm_re 704 115,095,930
cadforge_3d 14 35,785
cadforge_cq 3 9,347
multiturn 192 194,013
unified 758 148,433
vision_docmatix 1617 225,903
vision_forge 1232 68,406
plot2code 379 212,787
sophia_svg 23863 57,107,843
icons_forge 8490 3,043,056
re_synth_forge 6059 17,042,338
cad_synth_forge 4878 1,092,433
cad_evo_forge 11760 19,589,906
longctx_forge 30 49,735,767
blender_forge 3975 1,220,418
godot_forge 87 25,810
plot_forge 8021 3,914,630
godot_demos_real 248 221,945
godot_docs_real 2429 681,657
mpl_gallery_real 184 187,999
vision_forge2 7997 526,001
cq_examples_real 28 17,503

Pipeline

carry vetted v1.1 shards -> top-up fetch (rows+parquet bulk, offset-cursor resume) -> per-row gates (shape, non_latin, think-balance, dup, license, no_image) -> ONE global dedup index seeded with the carried corpus -> real gpt2 BPE per row -> per-slice pareto (strict) -> domain token-share caps -> deterministic interleave.

Quota drops (rebalance)

family rows dropped share before
reasoning_math 173173 0.2632
reasoning_sci 16299 0.1769
agentic_swe 506 0.1453
code_agent 194 0.1114
reverse_engineering 2 0.0688

Generator diversity (per source, post-pareto)

source distinct tasks top task top share
openmath_cot 6 math 55%
opensci 2 science 57%
openmath_tir 4 aops_c6_high_school_olympiads 54%
openr1_math 7 olympiads 73%
pyromind_short 2 tooluse 57%
toolmind 2 tooluse 54%
nemotron_swe 2310 swe 8%
swesmith 1685 swe 42%
coderforge 801 swe 29%
ultradata_code 3001 coding 3%
ultradata_general 2001 general 8%
ultradata_search 2001 search 0%
ollvm_re 2 re 100%
vision_docmatix 2 vision 88%
plot2code 2 vision 97%
icons_forge 9 tabler-outline 39%
re_synth_forge 2937 rle 2%
cad_synth_forge 23 plate_holes 12%
cad_evo_forge 1258 box_stack_pocket_boss 12%
longctx_forge 6 utility library header rebuild 37%
blender_forge 47 camera_rig 34%
godot_forge 20 tween_motion 20%
godot_demos_real 99 demo_2d_physics_tests 10%
godot_docs_real 1247 docs_class_openxrspatialcompo 1%
mpl_gallery_real 24 gallery_linebar 17%
vision_forge2 6 vf2_scatter_quads 17%

Single-artifact families carry a concision envelope: at equal correctness the pareto prune now drops bloated rows first, so the mixture teaches tight, dense answers rather than padded ones.

Files

  • train.jsonl.gz — the mixture (interleaved; shuffle again before packing)
  • manifest.jsonl.gz — per-row index: id, source, domain, turns, gpt2_tokens, tags
  • build/v2.0/shards/*.jsonl.gz — per-source provenance pools
  • images.tar.gz — PNGs referenced by vision_image/vision_plot rows (meta.images)

Tag vocabulary (filter, never re-split)

src:<source> · domain:<fine> (agentic_general, agentic_search, agentic_swe, agentic_tool, blender_3d, cad_3d, code_agent, general_sft, godot_dev, reasoning_math, reasoning_sci, reasoning_science, reverse_eng, reverse_engineering, style_house, vision_image, vision_plot, vision_svg) · turns:<mono|short|mid|deep|mega> · tok:<1k|1k-5k|5k-20k|20k-50k|50k+> · tools + toolcalls:<1-3|4-10|10+> · cot · multimodal · verified

Honest notes

  • Carried rows keep their v1.0/v1.1 vetting; meta.license is per-row.
  • Carry-only sources: agenticsft/demyagent (upstream license unverified), cadforge_3d (upstream dark since v1.0 — the 14 verified rows are all that exist).
  • Vision rows' text token share is small by nature: the pixels live in images.tar.gz, outside any text token count. Every forge-rendered vision answer is program-verified upstream.
  • Assistant turns carry trainable: true — mask non-assistant turns for loss.