eewer/swerebench-traces-raw-source-verification-enhanced-20260617

SWE-rebench Raw Source Verification Enhanced 20260617

This is a private raw source dataset for building refined mini-swe-agent SFT datasets. It is intentionally not tokenized and intentionally preserves source data plus metadata for downstream filtering, masking, weighting, and audit. Do not treat every row as a clean endpoint solve.

Download

The full dataset directory is uploaded as a single compressed archive:

hf download eewer/swerebench-traces-raw-source-verification-enhanced-20260617 \
  swerebench-traces-raw-source-verification-enhanced-20260617.tar.zst \
  --repo-type dataset --local-dir ./download

tar --zstd -xf ./download/swerebench-traces-raw-source-verification-enhanced-20260617.tar.zst

Archive details:

File	Size	SHA256
swerebench-traces-raw-source-verification-enhanced-20260617.tar.zst	2,027,391,851 bytes (1.888 GiB)	b0085a7b26539d592eac53274944d627c95826da40078db19799cf8a04dfb368

The archive expands to:

swerebench-traces-raw-source-verification-enhanced-20260617/
  data/*.jsonl.zst
  metadata/*
  manifest.json

The JSONL shards are already zstd-compressed. The outer tar.zst is provided so the complete source dataset can be downloaded as one file.

High-Level Counts

Metric	Count
Rows	41,784
Unique UUIDs	41,784
Unique tasks	28,780
Data shards	37
Metadata files	91
Source directory bytes before tar	2,991,954,690
Rows with >=90% assistant reasoning	40,836
Rows below 90% assistant reasoning	948
Rows with non-empty model patch	39,411
Rows with submit status	38,923
Rows with API calls	40,710
Rows with compaction original row ID	5,469
Rows with compaction model reasoning metadata	4,350
Synthetic verification/postprocessing rows	208

Rows By Language

Language	Rows
c	844
cpp	2,310
go	8,423
java	2,493
js	5,434
php	2,365
python	9,755
rust	4,175
ts	5,985

Rows By Pass Label

Passed	Rows
false	27,840
true	13,944

Important: verification recovery-prefix and stop rows are preserved as source data and may have pass-like inherited context. Use metadata such as verification_should_not_be_counted_as_passed and the audit sidecars before treating rows as endpoint-passed solves.

Rows By Teacher

Teacher	Rows
deepseek-v4-flash	1,803
deepseek-v4-pro	1,781
deepseek/deepseek-v4-flash	882
inclusionai/ring-2.6-1t	90
local qwen3.6-35b-a3b-fp8	1,819
xiaomi/mimo-v2.5	34,900
xiaomi/mimo-v2.5-pro	509

Rows By Difficulty

Difficulty	Rows
easy	15,138
hard	1,624
medium	25,022

Rows By Instruction Style

Instruction style	Rows
deepswe	8,457
mimo_clean_combined	110
mimo_clean_observation	10
mimo_clean_strict_baseline	10
mimo_clean_system	10
original	33,092
original_blocks	95

Rows By Source

Row source	Rows
audited_original_harness_c_cpp_topup	75
compaction_prefix_v2	2,333
compaction_prefix_v2_first_turn_reasoning_repaired	1,012
compaction_prefix_v2_prompt_and_first_turn_repaired	1,012
compaction_prefix_v2_prompt_and_first_turn_repaired_v5	1,112
existing_all_trajectories	3,526
generated_raw_trace	3,224
manifest_tsv	1,326
mimo_manifest_trace	10,406
other_source_exact_benchmark_strict_audit_trace	30
other_source_strict_audit_trace	292
post_260609_run_trace	16,602
post_parent_original_harness_c_cpp_addon	16
result_index	609
swesmith_c_cpp_clean_traces_checkpoint	63
synthetic_empty_submit_verification_stop	146

Verification Enhancement

This dataset extends:

/wbl-fast/usrs/ee/code-swe-data/runtime/hf_upload/swerebench-traces-raw-source-targeted-limitations-compaction-prompt-firstturn-repaired-v5-1000plus-20260617

with 208 verification/postprocessing rows in data/train-00036.jsonl.zst.

Verification family	Rows
synthetic_empty_patch_one_turn_recovery	56
synthetic_empty_submit_verification_stop	146
synthetic_patchtxt_verification_prototype	6

Recommendation metadata for the appended verification rows:

Signal	Count
verification_recommended_for_standard_sft=true	58
verification_recommended_for_standard_sft=false	150
verification_recommended_for_recovery_prefix_training=true	200
verification_recommended_for_recovery_prefix_training=false	2
verification_should_not_be_counted_as_passed=true	202
verification_should_not_be_counted_as_passed=false	6

Inherited-row verification scan summary from metadata/verification_enhanced_summary.json:

Signal	Rows
Inherited rows scanned	41,576
Natural patch verification	20,900
Missing patch verification	15,948
Missing non-empty patch verification	15,790
Missing empty patch verification	158
Natural empty-patch recovery produced non-empty patch	25

Synthetic-turn audit summary from metadata/verification_enhanced_synthetic_turn_audit.json:

Audit signal	Value
All synthetic rows audited in context	True
Synthetic verification reasoning unique strings	24
Synthetic verification max exact repetition	9
Empty-stop terminal reasoning unique strings	20
Empty-stop terminal max exact repetition	8
Generated recovery reasoning unique strings	56
Generated recovery reasoning max exact repetition	1

Recovery quality counts:

Quality	Rows
bad	1
high_confidence	29
needs_audit	1
plausible	25

Compaction Metadata

Compaction rows are included as raw source data. They are not filtered out. Use lineage fields to relate compacted descendants to original rows:

• metadata.compaction_original_row_id
• metadata.compaction_original_row_path
• metadata.prompt_repair_source_raw_compacted_uuid
• metadata.prompt_repair_source_firstturn_uuid
• metadata.recommended_for_compaction_training
• metadata.compaction_model_reasoning or metadata.compaction_model_reasoning_hash

Compaction recommendation counts in this full dataset:

recommended_for_compaction_training	Rows
false	24
true	3,112

Rows by compaction/source family:

Family	Rows
audited_original_harness_c_cpp_topup	75
compaction_prefix_v2	2,333
compaction_prefix_v2_first_turn_reasoning_repaired	1,012
compaction_prefix_v2_prompt_and_first_turn_repaired	1,012
compaction_prefix_v2_prompt_and_first_turn_repaired_v5	1,112
existing_all_trajectories	3,526
generated_raw_trace	3,224
manifest_tsv	1,326
mimo_manifest_trace	10,406
other_source_exact_benchmark_strict_audit_trace	30
other_source_strict_audit_trace	292
post_260609_run_trace	16,602
post_parent_original_harness_c_cpp_addon	16
result_index	609
swesmith_c_cpp_clean_traces_checkpoint	63
synthetic_empty_submit_verification_stop	146

Token And Size Signals

This source dataset is not tokenized. Some rows have token sidecar metadata from prior Qwen3-thinking audits, but not all rows do.

Signal	Value
Rows with Qwen3 token metadata in-row	154
Qwen3 tokens min	2226
Qwen3 tokens p50	16403
Qwen3 tokens p95	65222
Qwen3 tokens max	201532
Trajectory chars p50	231,217
Trajectory chars p95	929,756
Trajectory chars max	11,264,270,576
Model patch bytes p50	2,244
Model patch bytes p95	19,291
Model patch bytes max	2,665,338

Important Metadata Files

• manifest.json: build manifest for the full dataset lineage.
• metadata/full_dataset_upload_stats_20260617.json: full-shard stats used for this README.
• metadata/verification_enhanced_summary.json: verification-enhancement summary.
• metadata/verification_enhanced_row_signals.jsonl: row-level natural/missing/synthetic verification signals.
• metadata/verification_enhanced_synthetic_turn_audit.json and .jsonl: in-context audit of all appended synthetic turns.
• metadata/strict_quality_index.jsonl: strict-quality row metadata for generated additions where available.
• metadata/rejected_index.jsonl: rejected/metadata-only rows where available. Empty in this final lineage, while separate rejected sidecars are preserved for specific compaction batches.
• metadata/compaction_*: compaction attempts, request/response metadata, prompt-boundary audits, token audits, model reasoning, skip lists, and recommended UUID lists.
• metadata/parent_index.* and metadata/appended_index.*: inherited and early appended row indexes from prior source datasets.

Data Shape

Each data shard is JSONL compressed with zstd. Rows are raw source records for later SFT construction. Common top-level keys include:

compaction_model_reasoning, compaction_model_response, deepswe_prompt_augmentation, messages, metadata, model_patch, passed, percent_messages_with_reasoning, prompt, prompt_repair, reasoning, reward, task_id, teacher, tools, uuid

Metadata keys are intentionally rich and vary by row family. The most common metadata keys are recorded in metadata/full_dataset_upload_stats_20260617.json under metadata_keys_top.

Caveats For Downstream Filtering

• This is a raw source dataset, not a final clean SFT dataset.
• Some rows are failed, partial, recovery-prefix, compacted, or synthetic post-processing rows.
• Do not train every terminal submit turn equally without consulting metadata.
• Rows with verification_should_not_be_counted_as_passed=true should not be counted as clean endpoint passes.
• Use compaction lineage fields to avoid training both an original long row and its compacted derivative when that is undesirable.
• Use percent_messages_with_reasoning, patch size, trajectory size, submit-quality metadata, and verification metadata for filtering/weighting.

Provenance

Local source path at upload time:

/wbl-fast/usrs/ee/code-swe-data/runtime/hf_upload/swerebench-traces-raw-source-verification-enhanced-20260617

Build version: verification_enhanced_20260617_v1

Upload README generated at Unix time 1781732833.