← TinyGPT · docs · devlog · roadmap · speedup
source: docs/dataset-inventory.md · view on GitHub ↗

Dataset inventory

Live snapshot of ~/.cache/tinygpt/datasets/ (not checked into the repo). Last refreshed 2026-06-17. Sizes are directory totals; row counts are from decoded JSONLs where available, otherwise marked as parquet-on-disk.

Specialist-track training corpora

DatasetPathSizeDecoded JSONLUseStatus
NousResearch/hermes-function-calling-v1NousResearch/113 MBhermes-fc.jsonl (49.8 MB)A1 tool-caller SFT (Apache 2.0)✅ decoded
Locutusque/function-calling-chatmlLocutusque/data/102 MBfunction-calling-chatml.jsonl (333 MB)A1 tool-caller SFT✅ decoded
Salesforce/xlam-function-calling-60kSalesforce/xlam-function-calling-60k/xlam_function_calling_60k.json91.7 MB~60K rowsA1 tool-caller SFT (gated)✅ pulled 2026-06-17 (token + license click)
yahma/alpaca-cleanedyahma/82 MBGeneral-instruction SFT baseline✅ on disk
Intel/orca_dpo_pairsIntel/67 MBDPO preference data✅ on disk
argilla/ultrafeedback-binarized-preferences-cleanedargilla/137 MBultrafeedback.jsonlDPO preference data✅ decoded
meta-math/MetaMathQAmeta-math/663 MBMath specialist SFT✅ on disk
iamtarun/python_code_instructions_18k_alpacaiamtarun/data/11 MBpython-code-instr.jsonl (18 612 rows)Code specialist SFT✅ decoded
bigcode/the-stack-smolbigcode/the-stack-smol/data/{c,c++,go,java,javascript,python,rust,typescript}/data.json~850 MB8 langs (of 30 in repo)Code specialist pretrain✅ pulled 2026-06-17 (token + license click)
HuggingFaceFW/fineweb-eduHuggingFaceFW/ + fineweb-edu.txt (230 MB)2.2 GBfineweb-edu.txt (50K-row sample)Pretrain quality baseline✅ sampled

ScaleDown (B25) corpora — pulled 2026-06-17

DatasetPathSizeUse
microsoft/ms_marco v1.1microsoft/ms_marco/v1.1/207 MB(query, doc, answer) triplets for context-compression training. Test 19.5 MB + train 167 MB + val 20 MB. v2.1 (7 train shards) not pulled — would add ~1 GB if needed.
google-research-datasets/natural_questionsgoogle-research-datasets/natural_questions/default/375 MBNQ training data subset (2 of 287 shards). Full pull would be multi-GB; current subset bounds disk.

Eval splits (D5)

DatasetPathRowsUse
openai/gsm8k (main)openai/gsm8k/main/1 319 test / 7 473 train (parquet)Math reasoning eval (E3 + E4)
HuggingFaceH4/MATH-500HuggingFaceH4/MATH-500/test.jsonl500Canonical MATH eval (subject, level, problem, solution, answer)
openai/openai_humanevalopenai/openai_humaneval/openai_humaneval/164 (parquet)Code-gen eval (E5)
google-research-datasets/mbpp (full)google-research-datasets/mbpp/full/974 (parquet, prompt+test+train+val)Code-gen eval (E5)
princeton-nlp/SWE-bench_Verifiedprinceton-nlp/data/500 (parquet)Code repair eval (later A1)

Indic / multilingual

DatasetPathSizeUse
google/IndicGenBench_xquad_ingoogle/IndicGenBench_xquad_in/45 MBA5 — Indic XQuAD splits across as/bn/gu/en + more
ai4bharat/MILUai4bharat/MILU/0 BA5 — MILU is an lm-eval-harness task, not raw rows; harness code is at _external/MILU/

Pace planner data (in-repo specialist work)

FileSizeUse
pace-prompts.jsonl / pace-prompts-v2.jsonl / pace-prompts-v3.jsonl4 KB → 39 KB → 12 KBPlanner prompt corpus (versioned)
pace-labeled.jsonl / pace-labeled-v2.jsonl / pace-labeled-v3.jsonl13 KB → 116 KB → 25 KBLabeled planner intents
pace-sft-v2.jsonl / pace-sft-v3.jsonl36 KB / 14 KBSFT pairs for planner
clarify-seeds.jsonl / clarify-train-v1.jsonl / clarify-dpo-v1.jsonl15 KB / 15 KB / 76 KBClarify-action seed + train + DPO

External evaluators (source code, not data)

_external/ holds checked-out harness repos that the tinygpt eval-* subcommands shell out to:

PathRole
_external/gorilla-bfcl/BFCL harness — invoked by tinygpt eval-bfcl (E1)
_external/tau-bench/τ-bench harness — invoked by tinygpt eval-tau-bench (E2)
_external/MILU/lm-eval-harness MILU task config

Outstanding gaps

  1. MS-MARCO v2.1 — only v1.1 is pulled; v2.1 has 7 train shards (~1 GB) if B25 needs additional training data.
  2. Natural Questions — only 2/287 train shards; pull more if B25 needs broader coverage.
  3. the-stack-smol additional langs — 8/30 pulled (c, c++, go, java, javascript, python, rust, typescript). Add more (haskell, lua, ruby, scala, kotlin, etc.) if the code specialist needs broader exposure.

All other items in PLAN.md Tier D are satisfied.