id string | tags list | messages list | meta unknown | approx_tokens int64 | gpt2_tokens int64 | meta_extra unknown | tools list | images list |
|---|---|---|---|---|---|---|---|---|
zai_overflow:afdab0eff4d54e68 | [
"style",
"concurrency",
"token-efficient",
"hard-verified",
"domain:style_house",
"turns:mono",
"tok:<1k"
] | [
{
"role": "user",
"content": "Draft a Slack message reply declining a 'quick sync meeting' request that could turn into a 30-minute discussion. Compare alternatives in a table.",
"trainable": false
},
{
"role": "assistant",
"content": "Would prefer not to schedule another meeting. Here are b... | {
"task": "style",
"difficulty": null,
"lang": null,
"domain": "style_house",
"mode": "batch",
"score": null,
"source": "unified",
"turns": 2
} | 115 | 108 | null | null | null |
ens_0d27f82a9307 | [
"domain:reasoning_science",
"turns:mono",
"tok:5k-20k",
"cot"
] | [
{
"role": "user",
"content": "Solve the following problem. Make sure to put the answer (and only answer) inside \\boxed{}.\n\nIn Horeau's method, a racemic sample of chiral secondary alcohol A is esterified with enantiomerically pure (R)-2-phenylbutanoic acid. After partial reaction, the unreacted alcohol i... | {
"source": "opensci",
"turns": 2,
"domain": "reasoning_science",
"task": "science",
"license": "cc-by-4.0"
} | 13,042 | 17,553 | {
"cot": "inline"
} | null | null |
ens_4f0f2a73f5a7 | [
"turns:mid",
"tok:1k-5k",
"src:re_synth_forge",
"domain:reverse_engineering"
] | [
{
"role": "user",
"trainable": false,
"content": "Here is x86-64 disassembly (objdump -d -M intel, no raw bytes) of a function compiled with GCC -O2 from a stripped binary. Reconstruct readable, compilable C that preserves the exact behavior. Explain the key control-flow decisions as you go.\n\n```asm\n... | {
"source": "re_synth_forge",
"domain": "reverse_engineering",
"license": "cc-by-4.0",
"task": "fixedpoint+parseip+vm",
"turns": 6
} | 1,111 | 2,459 | null | null | null |
ens_84c16f63d49e | [
"turns:short",
"tok:5k-20k",
"cot",
"src:openmath_cot",
"domain:reasoning_math"
] | [
{
"role": "user",
"content": "Find the closed form of the sum $\\left\\lceil \\sqrt{1} \\right\\rceil + \\left\\lceil \\sqrt{2} \\right\\rceil + \\cdots + \\left\\lceil \\sqrt{n} \\right\\rceil$.",
"trainable": false
},
{
"role": "assistant",
"content": "<think>\nOkay, so I need to find a cl... | {
"source": "openmath_cot",
"domain": "reasoning_math",
"license": "cc-by-4.0",
"task": "aops_c6_high_school_olympiads",
"turns": 2
} | 6,208 | 11,060 | null | null | null |
ens_4b0bae6e23ce | [
"domain:agentic_tool",
"turns:deep",
"tok:5k-20k",
"tools",
"toolcalls:4-10"
] | [
{
"role": "system",
"content": "<instructions>\nYou are a customer service agent that helps the user according to the <policy> provided below.\n\nCRITICAL RULES (you MUST follow these):\n1. In each turn, you can ONLY do ONE of these:\n - Send a message to the user, OR\n - Make exactly ONE tool call\n2. ... | {
"source": "ultradata_tooluse",
"turns": 22,
"economy": {
"tokens": 5667,
"calls": 8,
"waste_tokens": 0
},
"domain": "agentic_tool",
"task": "tooluse",
"license": "apache-2.0"
} | 5,674 | 6,692 | null | [
{
"type": "function",
"function": {
"name": "book_reservation",
"description": "Book a reservation.",
"parameters": {
"$defs": {
"FlightInfo": {
"properties": {
"flight_number": {
"description": "Flight number, such as 'HAT001'.",... | null |
ens_6d380743aebc | ["turns:mega","tok:20k-50k","tools","toolcalls:10+","tools","src:ultradata_code","domain:code_agent"(...TRUNCATED) | [{"role":"system","content":"You are Codex, based on GPT-5. You are running as a coding agent in the(...TRUNCATED) | {"source":"ultradata_code","domain":"code_agent","license":"apache-2.0","task":"Code_Agent_001478","(...TRUNCATED) | 45,457 | 76,107 | null | null | null |
ens_3079c6d15e5a | [
"domain:vision_plot",
"turns:mono",
"tok:<1k"
] | [{"role":"user","content":"Write Python matplotlib code that generates a figure matching this descri(...TRUNCATED) | {
"source": "plot2code",
"turns": 2,
"domain": "vision_plot",
"task": "vision",
"license": "apache-2.0"
} | 471 | 569 | {"plot_url":"https://matplotlib.org/stable/gallery/lines_bars_and_markers/scatter_masked.html#sphx-g(...TRUNCATED) | null | null |
ens_83005d904afe | [
"turns:short",
"tok:1k",
"src:cad_synth_forge",
"domain:cad_3d"
] | [{"role":"user","content":"Write CadQuery code that produces the parametric part described below. Th(...TRUNCATED) | {
"source": "cad_synth_forge",
"domain": "cad_3d",
"license": "cc-by-4.0",
"task": "plate_holes",
"turns": 2
} | 157 | 209 | null | null | null |
ens_014c24ad7a56 | ["turns:mega","tok:5k-20k","tools","toolcalls:10+","tools","src:ultradata_general","domain:agentic_g(...TRUNCATED) | [{"role":"system","content":"You are a personal assistant running inside OpenClaw.\n## Tooling\nTool(...TRUNCATED) | {"source":"ultradata_general","domain":"general_sft","license":"apache-2.0","task":"General_Agent_00(...TRUNCATED) | 18,695 | 25,306 | null | null | null |
ens_3ae3783e031e | [
"domain:vision_image",
"turns:mono",
"tok:<1k",
"multimodal"
] | [{"role":"user","content":"What type of content does the website mentioned in the text offer?","trai(...TRUNCATED) | {"source":"vision_docmatix","turns":2,"domain":"vision_image","task":"vision","license":"apache-2.0"(...TRUNCATED) | 46 | 35 | {
"modality": "image+text"
} | null | [
"images/vision_80701df9c6.png"
] |
End of preview. Expand in Data Studio
filterme-sft · v1.7
Current release:
v1.7— the repo root always holds the latest release: finals (train.jsonl.gz,manifest.jsonl.gz), per-source provenance pools (shards/), build state and the QA report. One clean base — no historical build trees.
| rows | real gpt2 tokens | sources | pareto pruned | QA |
|---|---|---|---|---|
| 165,545 | 2,004,534,369 | 30 | 13,622 | all gates green |
One pareto-pruned, token-rebalanced SFT mixture built by sft-forge: ingest → tag → gate → ONE global dedup → strict per-slice pareto (quality vs real gpt2 token cost) → domain token-share caps → deterministic interleave. Every row carries fine-grained tags and a true BPE token count — filter, never re-split. The name is the doctrine: filtered, not scraped.
Version history
| version | rows | gpt2 tokens | sources | highlights |
|---|---|---|---|---|
| v0.3 | 1 | 123 | 1 | smoke fixture. Briefly the public card by accident (a stale upload_folder clobber); retired for good in this push. |
| v1.0 | — | — | — | first ensemble: multi-source ingest, ONE global dedup index, per-slice strict pareto. |
| v1.1 | — | — | — | scale-up + license-purity gates (Apache/MIT/CC-BY-4.0/ISC) on every forge lane. |
| v1.2 | 149,358 | 1,683,409,624 | 24 | first full publication audit (0 diffs vs card); incremental per-source shard pushes. |
| v1.3 | 128,008 | ~1,452,443,000 | 29 | lanes: sophia_svg, icons_forge, re_synth_forge, cad_synth_forge, longctx_forge. |
| v1.4 | 130,797 | 1,352,839,219 | 29 | caps reshuffle + generator diversity v2 (RE share 0.1% → 13.7%); concision envelopes. |
| v1.5 | 139,707 | 1,452,971,107 | 30 | cad_evo exec-verified CAD evolution sessions; longctx monster recovery; quota_keep rewrite (normalized caps, tolerance exit, input-robust). |
| v1.6 | 137,956 | 1,462,487,838 | 30 | lingctx growth (128 mega sessions, 199.8M gpt2 kept) + cad_evo tighter-tolerance replacement (0.25% / 0.10 mm verify band, 4-7-step arcs, 41% repair loops, 2,632 kept); caps rebalanced to keep the monsters in-mix. |
| v1.7 ← current | 165,545 | 2,004,534,369 | 30 | the 2B release on the clean filterme-sft base: caps re-solved to convert the pool's unused 1.5B tokens (+542M vs v1.6); longctx lane grows again (241 mega sessions in-final, 374M gpt2); cad_evo +1,620 tight-band sessions; fresh repo, no legacy build trees. |
What's new in v1.7
- the 2B conversion — the v1.6 pool held ~3.1B real gpt2 tokens but the releases only consumed 1.46B (quota was cap-constrained, not pool-constrained). v1.7 re-solved the caps against the exact post-pareto pool (
scripts/design_caps_v17.py: bisection on the elastic-family scale, validated with the productionquota_keep) and converted the headroom: 2,004,534,369 real gpt2 tokens (+542M vs v1.6) with every domain family still inside its tolerance band (max violation 1.25pp). - longctx grows again — +150 fresh mega sessions (seed base 40000+) join the 301 carried; 241 survive in-final (374M gpt2 tokens) — hundred-class turn counts over REAL executed state: tools run, files mutate, every citation checked against prior tool output. These are the rows that teach long-horizon coherence.
- cad_evo tight-band growth — +1,620 new evolution sessions at the 0.25% / 0.10 mm verify band (41% real repair loops, 204+ arc families, 12/12 sampled re-exec inside the band). CAD and 3D keep their own slice of the mix with exec-verified only.
- clean base, new name — this release lives at IO-AIR/filterme-sft: a fresh repo with root-level files only (
shards/,state/, cards, finals). Nobuild/<version>/history trees, no smoke leftovers, no stale cards — the accumulated 241-file legacy of sft-forge-ensemble stays behind.
Token mix by domain (real gpt2, kept rows)
| domain | gpt2 tokens | share | |
|---|---|---|---|
| reasoning_math | 381,589,384 | 19.04% | █████··················· |
| agentic_swe | 359,611,343 | 17.94% | ████···················· |
| reasoning_sci | 294,581,048 | 14.70% | ████···················· |
| reverse_engineering | 284,560,356 | 14.20% | ███····················· |
| code_agent | 220,638,595 | 11.01% | ███····················· |
| agentic_tool | 219,243,982 | 10.94% | ███····················· |
| general_sft | 198,958,112 | 9.93% | ██······················ |
| vision_svg | 40,415,056 | 2.02% | ························ |
| cad_3d | 4,103,436 | 0.21% | ························ |
| style_house | 325,961 | 0.02% | ························ |
| vision_image | 294,309 | 0.01% | ························ |
| vision_plot | 212,787 | 0.01% | ························ |
Deep agentic trajectories dominate raw token share by design — they are the rows small models cannot synthesize. Rebalance a training mix by dropping or down-weighting tagged rows via the tag filter, never by re-splitting.
Sources (30)
Kept rows/tokens sum exactly to the release totals; pool rows are the provenance shard sizes (build/<version>/shards/) before global dedup, pareto and quota.
| source | kept rows | kept gpt2 | pool rows |
|---|---|---|---|
| longctx_forge | 241 | 386,971,361 | 451 |
| openmath_cot | 47,510 | 381,589,384 | 66,787 |
| opensci | 49,751 | 294,581,048 | 65,779 |
| toolmind | 13,139 | 171,025,036 | 13,351 |
| ultradata_code | 2,764 | 163,636,053 | 4,200 |
| ollvm_re | 704 | 115,095,930 | 704 |
| nemotron_swe | 1,694 | 95,892,622 | 4,702 |
| swesmith | 2,880 | 94,358,439 | 5,174 |
| coderforge | 948 | 78,556,491 | 2,184 |
| pyromind_long | 987 | 58,427,013 | 1,000 |
| ultradata_search | 1,306 | 39,506,375 | 2,576 |
| sophia_svg | 15,948 | 38,153,242 | 15,948 |
| ultradata_general | 1,338 | 33,996,539 | 2,692 |
| pyromind_short | 1,846 | 14,409,738 | 1,932 |
| re_synth_forge | 5,100 | 12,075,013 | 6,059 |
| ultradata_tooluse | 1,226 | 8,740,063 | 2,095 |
| agenticsft | 500 | 8,115,596 | 500 |
| cad_evo_forge | 3,593 | 3,389,137 | 5,820 |
| icons_forge | 6,354 | 2,261,814 | 6,354 |
| nemotron_cuda | 88 | 847,831 | 2,276 |
| mcp_atlas | 31 | 729,868 | 462 |
| cad_synth_forge | 3,204 | 709,604 | 3,204 |
| demyagent | 228 | 628,420 | 500 |
| vision_docmatix | 1,617 | 225,903 | 1,617 |
| plot2code | 379 | 212,787 | 379 |
| multiturn | 177 | 177,528 | 192 |
| unified | 758 | 148,433 | 758 |
| vision_forge | 1,232 | 68,406 | 1,232 |
| cadforge_cq | 1 | 2,667 | 3 |
| cadforge_3d | 1 | 2,028 | 14 |
| total | 165,545 | 2,004,534,369 | 218,945 |
Pipeline
- carry vetted shards from the previous release (byte-faithful, verified against
pushed_sources.json) - top-up fetch — rows + parquet bulk with offset-cursor resume
- per-row gates — shape, non-Latin, think-balance, dup, license, no_image
- ONE global dedup index seeded with the carried corpus (Jaccard 0.65)
- real gpt2 BPE count per row — no estimates anywhere
- strict per-slice pareto — quality vs token cost; equal-correctness rows drop the bloated one first (concision envelopes)
- domain token-share caps — normalized-caps cascade, tolerance exit (
quota_keep; seequota_metain ENSEMBLE_CARD.json) - deterministic interleave — same seed, same bytes, reproducible
QA gates (all green)
| gate | scope | result |
|---|---|---|
| JSON parse | all 165,545 rows | 0 bad |
| forbidden keys / bad ids | all rows | 0 / 0 |
| SVG XML parse + viewbox + forbidden attrs | 22,302 vision_svg | 0 fail · 0 missing · 0 forbidden |
| asm presence + length floor + session depth | 5,101 RE | 0 no-asm · 0 thin · 0 short |
cadquery + result contract |
3,204 CAD | 0 no-cq · 0 no-result |
| longctx turn/size floors | 241 sessions | 0 short · 0 small |
| domain caps (normalized, tolerance exit) | 12 families | max violation 0.0114 ≤ 0.0125 in 4 rounds |
License ledger (per-row meta.license)
| license | rows |
|---|---|
| cc-by-4.0 | 112,925 |
| apache-2.0 | 41,533 |
| mit | 8,952 |
| isc | 925 |
| cc-by-4.0 (per-row) | 223 |
| apache-2.0/mit (per-row) | 52 |
| house-generated (no upstream license) | 935 |
| total | 165,545 |
License doctrine: only Apache/MIT/CC-BY-4.0/ISC sources are adopted; sources governed by individual model providers' terms are excluded outright. Rows keep their per-row license in meta.license; carried rows keep their original vetting.
Files
train.jsonl.gz— the mixture (interleaved; shuffle again before packing)manifest.jsonl.gz— per-row index: id, source, domain, turns, gpt2_tokens, tagsshards/*.jsonl.gz— per-source provenance poolsstate/— progress, pushed_sources, qa_reportimages.tar.gz— PNGs referenced by vision rows (meta.images)
Format
{"id", "tags", "messages" (assistant turns loss-mask-ready via `trainable`), "approx_tokens", "gpt2_tokens", "meta"}
Tag vocabulary (filter, never re-split)
src:<source> · domain:<fine> (agentic_general, agentic_search, agentic_swe, agentic_tool, cad_3d, code_agent, general_sft, reasoning_math, reasoning_sci, 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
Usage
from datasets import load_dataset
ds = load_dataset("IO-AIR/filterme-sft", split="train") # or streaming=True
# row: {id, tags, messages, trainable, approx_tokens, gpt2_tokens, meta}
# loss masking: train only where trainable is True (assistant turns)
# svg_rows = [r for r in ds if "domain:vision_svg" in r["tags"]]
# mega = [r for r in ds if "turns:mega" in r["tags"]]
Related
- IO-AIR/arenamax — companion human-preference dataset (arena winners + rating-best-of-N), published separately with SFT and DPO configs.
Honest notes
- Carried rows keep their earlier-release vetting;
meta.licenseis per-row. - Legacy note: ~24k carried
ens_*rows pre-date the per-messagetrainableflag — for those, mask by role (train on assistant turns). All v1.3+ rows carry explicittrainablebooleans. Optional keys beyond the base schema:meta_extra(provenance detail),tools(tool-call rows),images(vision rows). - 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. - Every generated CAD row was volume/bbox-verified against an analytic exact answer; every RE row carries real asm from gcc/objdump.
- Downloads last month
- -