Hugging Face's logo

Datasets:
Fzz1
/
tb-training-mix-agentpick

Tasks:
Reinforcement Learning
Modalities:
Tabular
Text
Formats:
parquet
Languages:
English
Size:
1K - 10K
Tags:
terminal-bench
agentic-rl
verifier-audit
Libraries:
Datasets
pandas
Polars
Dataset card Data Studio Files
xet
 Community
1
tb-training-mix-agentpick / data

Commit History

Today

  1. Raise est_disk_mb 1024->20480 (1->20 GB) for 3 CalibForge clean-fast tasks that hit 'No space left' in 1030c training (torchtitan#309); only calibforge_clean_fast, 3 rows x est_disk_mb, nothing else changed

    Fzz1
    Verified signature
    bcb9cb9

  2. Drop B10 games_20260611_112034_001 and one CalibForge security row (B19) from calibforge_clean + calibforge_clean_fast (user decision, torchtitan#309); card counts + BUILD_RECORD correction; tars and every other file byte-identical

    Fzz1
    Verified signature
    f39167d

4 days ago

  1. Remove tw_14495 (Terminus-incompatible: jessie tmux 1.9 lacks pid/socket): drop from clean + clean_fast; card counts + BUILD_RECORD correction; data files otherwise unchanged

    Fzz1
    Verified signature
    78c6274

  2. Resource correction v4: 3 ids -> max 8192 MB RAM / 10240 MB disk, optimization_20260606_172622_001 -> req_cpus 4; 4 split(s); nothing else changed

    Fzz1
    Verified signature
    c180c10

5 days ago

  1. Resource correction v3: 15 run-1026b OOM/memory/disk ids -> max RAM 8192 MB / max disk 10240 MB (CPU unchanged); 4 split(s); nothing else changed

    Fzz1
    Verified signature
    be2abb2

6 days ago

  1. Resource correction v2: tw_17238, tw_17688, system-administration_20260605_015435_004 set to a uniform req_cpus=4 / req_memory_mb=8192 / est_disk_mb=4096 (disk raised to 4096 for build headroom); 4 split(s); nothing else changed

    Fzz1
    Verified signature
    84fc4b4

  2. Resource correction: tw_17238, tw_17688, system-administration_20260605_015435_004 req_cpus/req_memory_mb/est_disk_mb from the measured 4/8/10 model-rollout peak (1/1/1 died 'failed to resolve container IP'); 4 split(s); nothing else changed

    Fzz1
    Verified signature
    61dd40c

8 days ago

  1. Five rows to repair after a post-publish verifier audit 2026-10-24: clean 341 -> 340, repair 170 -> 171, calibforge_clean 401 -> 397, calibforge_repair 226 -> 230; the fast copies drop them

    Fzz1
    Verified signature
    7dee55f

16 days ago

  1. clean_fast / calibforge_clean_fast: tiered gap caps (3 s, 30 s) for the rows cap 0 left unconverted

    Fzz1
    Verified signature
    a539d6a

  2. Add clean_fast / calibforge_clean_fast: offset-compressed copies of clean / calibforge_clean

    Fzz1
    Verified signature
    475098c

17 days ago

  1. CalibForge: machine-learning_20260604_004411_001 calibforge_clean -> calibforge_repair (post-publish verifier audit 2026-10-15: INSTR-VERIFIER-MISMATCH)

    Fzz1
    Verified signature
    f7660a6

  1. CalibForge: the 3-rollout rule and the added ids (calibforge_clean 402, calibforge_repair 225; canary off, siblings repair)⁠

    Fzz1
    Verified signature
    ff04f19

  2. Xiaomi MiMo: the 4 opensource-code tasks without a solution move to xiaomi_repair (xiaomi_clean 4, xiaomi_repair 8, xiaomi_swe 20)⁠

    Fzz1
    Verified signature
    c4750c6

  3. Add the Xiaomi MiMo splits: xiaomi_clean (4), xiaomi_repair (4), xiaomi_swe (24)⁠

    Fzz1
    Verified signature
    a7216c5

  1. Add the Terminal-Lego splits: terminallego_clean (43 rows) and terminallego_repair (13 rows)⁠

    Fzz1
    Verified signature
    730bf3e

  1. tb-training-mix-agentpick: CalibForge in two splits of its own (calibforge_clean 240, calibforge_repair 51); clean and repair unchanged; card: CalibForge section and each re-pinned row's Debian release

    Fzz1
    Verified signature
    6e3e6c8

  2. tb-training-mix-agentpick: the published 440 rows plus TerminalWorld on v2.1.1 (clean 50, repair 21); CalibForge is not included in this version

    Fzz1
    Verified signature
    8ac3dbe

  3. tb-training-mix-agentpick: add CalibForge taken whole (289 shipped: clean 238, repair 51) beside the 440 published rows; TerminalWorld still deferred

    Fzz1
    Verified signature
    6eda006

  1. tb-training-mix-agentpick: req_* takes the maximum across measured sources (our newest, the replay, the upstream agent rollout); 21 SWE rows get more RAM

    Fzz1
    Verified signature
    9858636

  2. tb-training-mix-agentpick: directly consumable by the training side -- shard, environment/Dockerfile in every TMAX package, req_* filled from measured peaks; 440 rows, TerminalWorld follows later

    Fzz1
    Verified signature
    ea9aa0b

  1. tb-training-mix-agentpick: add SWE-Smith 37 + SWE-Rebench 61 -- 440 rows, clean 291 / repair 149; card describes the agent re-rank; TerminalWorld follows in a later commit

    Fzz1
    Verified signature
    e23cf1d

  2. tb-training-mix-agentpick: agent-reranked TB2.1 top-10, clean 195 / repair 147

    Fzz1
    Verified signature
    e54a0dc