Hugging Face's logo

Datasets:
Changyeli03
/
hle-flowbench-experiments-20260829

Modalities:
Tabular
Text
Formats:
csv
Languages:
English
Chinese
Size:
< 1K
Tags:
flowbench
hle
agent-evaluation
flame-chase
ralph-loop
experiment-traces
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
Dataset Viewer
Auto-converted to Parquet Duplicate
archive
stringlengths
18
45
compressed_bytes
int64
805
879M
members
int64
5
345k
regular_files
int64
3
180k
symlinks
int64
0
77.9k
zstd_valid
bool
1 class
tar_stream_valid
bool
1 class
banned_name_hits
int64
0
0
known_secret_hits
int64
0
0
heuristic_token_hits
int64
0
0
status
stringclasses
1 value
runs/additional-5x-fixed-2h.tar.zst
103,337,731
209,698
96,079
12,744
true
true
0
0
0
pass
runs/api-error-backfill-2h.tar.zst
2,123,013
3,058
1,786
80
true
true
0
0
0
pass
runs/archive.tar.zst
313,872,209
74,335
38,752
7,797
true
true
0
0
0
pass
runs/fixed-2h.tar.zst
878,676,616
345,169
179,914
77,930
true
true
0
0
0
pass
runs/glm-empty-backfill-2h.tar.zst
805
5
3
0
true
true
0
0
0
pass
runs/gpt-kimi-fixed-2h.tar.zst
332,490,041
221,598
88,923
23,978
true
true
0
0
0
pass
runs/humanize-goal-6h.tar.zst
51,258,083
24,047
12,288
1,119
true
true
0
0
0
pass
runs/k3-c60-reruns-2h.tar.zst
217,482,223
287,308
118,909
18,110
true
true
0
0
0
pass
runs/opus-k3-c60-reruns-2h.tar.zst
73,287,559
197,217
86,576
21,281
true
true
0
0
0
pass
runs/smoke.tar.zst
713,880
1,435
790
45
true
true
0
0
0
pass
sessions/host-hle-orchestration.jsonl.tar.zst
153,958,883
105
105
0
true
true
0
0
0
pass

HLE FlowBench experiment archive

Private migration snapshot of the local HLE text-only 100-question research program through 2026-09-01. It preserves the formal and smoke runs, per-question Codex/Claude/Kimi sessions, workflow attempts and metrics, evaluator state, scores, monitoring, experiment controllers, reports, analyses, the paused-run migration package, source Git bundles, and HLE-related host orchestration sessions.

The current Chinese experiment status, validity qualifications, and executable rerun priorities are in STATUS_AND_REQUIRED_RERUNS.md. The original authoritative local summary is mirrored at reports/HLE_AOPT_FINAL_SUMMARY.md.

Formal results, including completed K3 c60 reruns

Workflow Configuration Completion date (UTC) Final average Status
Flame Chase Opus 5 + GPT-5.6 Sol 2026-08-22 67.00% Complete
Flame Chase Opus 5 + Kimi K3 (original) 2026-08-27 67.00% Superseded: K3 90.92% terminal 429
Ralph Loop Opus 5 2026-08-28 67.00% Complete after API-only backfill
Humanize /goal Opus 5 2026-08-28 66.67% Complete; five capped goal failures count wrong
Flame Chase Opus 5 + Kimi K3 (c60 v2) 2026-09-01 66.00% Complete; 0 terminal 429, degraded by 1,409 terminal balance-402 sessions
Flame Chase GPT-5.6 Sol + Kimi K3 (c60 v2) 2026-08-31 65.00% Complete; 0 terminal 429, degraded by 2,065 terminal balance-402 sessions
RLCR Opus 5 + GPT-5.6 Sol 2026-08-22 64.67% Complete
Flame Chase GPT-5.6 Sol + Kimi K3 (original) 2026-08-24 64.33% Superseded: K3 14.60% terminal 429
Ralph Loop GPT-5.6 Sol 2026-08-23 61.33% Complete
Ralph Loop Kimi K3 2026-08-28 61.33% Complete after API-only backfill
Flame Chase Kimi K3 + Qwen38Max (c60 v2) 2026-08-31 60.33% Complete; 0 terminal 429, degraded by 5,062 terminal balance-402 sessions
RLAR GPT-5.6 Sol + GPT-5.6 Sol 2026-08-22 60.33% Complete
Humanize /goal GPT-5.6 Sol 2026-08-28 59.33% Complete
Flame Chase Kimi K3 + Qwen38Max (original) 2026-08-28 56.33% Superseded: K3 31.78% terminal 429
Ralph Loop Qwen38Max 2026-08-27 49.00% Complete
Ralph Loop DeepSeek V4 Pro 2026-08-27 47.00% Complete
Ralph Loop GLM 5.3 2026-08-28 43.00% Complete after API-only backfill

The machine-readable equivalents are inventory/formal_results.jsonl, inventory/formal_cells.jsonl, inventory/k3_session_health.json, inventory/k3_c60_rerun_summary.json, and inventory/required_reruns.jsonl. The c60 score and provider-health qualification is explained in reports/HLE_K3_C60_FINAL_RESULTS.md.

Coverage

  • 17 versioned formal model/workflow groups, 51 cells, and 5,100 locked question results, including all three completed K3 concurrency-60 reruns
  • 39 API-error-only supplement events, with immutable originals and merged corrected evaluator copies
  • six native Humanize /goal formal cells, plus the failed pre-fix attempt and smoke/regression runs
  • all retained run families: historical archive, fixed two-hour runs, GPT/Kimi runs, five added groups, /goal, API backfill, GLM backfill scaffold, smoke runs, and both K3 concurrency-60 rerun families
  • per-question Codex JSONL, Claude project JSONL, Kimi event journals and Kimi wire/session state retained inside the run archives
  • raw local inventory: 80,435 Codex session JSONL, 67,904 Claude project transcripts, 56,518 Kimi event journals, and 56,515 Kimi wire JSONL
  • HLE-matching host Codex orchestration transcripts retained separately
  • paused 2026-08-19 portable migration package, including its exact Docker image bundle and verified checkpoint

Layout

  • runs/*.tar.zst: one full archive per local run family; the recent additional-run monitoring snapshots and API-backfill merged evaluator copies are also mirrored as browsable directories so current report links work
  • sessions/host-hle-orchestration.jsonl.tar.zst: host orchestration sessions
  • reports/, experiments/, analysis/, monitoring/: browsable local files
  • migration/2026-08-19-paused/: original verified portable checkpoint package
  • sources/git/: Git bundles and provenance for FlowBench, Flowverse, and the Humanize /goal fix branch
  • inventory/: scores, cells, session health, archive validation, file sizes, SHA-256 checksums, and security-scan results
  • configs/required-reruns.yaml: machine-readable P0/P1 plan and its final execution/quality status; it is not an auto-launch file
  • MIGRATION.md: restore and verification instructions
  • EXCLUSIONS.md: credential/cache exclusion boundary

Download and verify 

hf download Changyeli03/hle-flowbench-experiments-20260829 \
  --type dataset \
  --local-dir hle-flowbench-experiments-20260829
cd hle-flowbench-experiments-20260829
sha256sum -c inventory/sha256sums.txt

The repository is private because complete research transcripts contain model reasoning, prompts, task content, and machine-local paths. Authentication files, tokens, environment snapshots, SQLite state databases, and regenerable caches are excluded.
Downloads last month
185
Number of rows:
11
Total file size:
6.54 GB