The remembered "3392" high-quality count appears to come from estimating 848 steps * global_batch_size 16 / 4 epochs. The audited on-disk dataset contains 3384 unique high-quality rows.

Files

data/training_dedup_clean_5405.parquet
- Recommended train-ready deduplicated dataset.
- 5405 rows.
- Contains the cleaned combination that was used for the current Qwen3.5-9B training line.
data/consolidated_three_sources_full_5944.parquet
- Full provenance/audit view.
- 5944 rows.
- Includes all 832 original GPT-5.4-NANO legacy rows, even when a query overlaps newer fixed-thinking rows.
data/summary.json
- Machine-readable source paths, counts, and notes.
src/build_consolidated_dataset.py
- Rebuild script.

The JSONL files in data/ are local convenience copies generated by the builder; the dataset repo should use the Parquet files as the primary artifacts.

Columns

query_id: Original question/run identifier.
origin_type: Provenance category. This is the main origin identifier.
source_view: Whether the row belongs to the train-ready deduplicated view or the full provenance view.
process_bucket: Original process bucket from the conversion pipeline.
source: Source label from the upstream metadata.
gold_answer: Ground-truth answer.
predicted_answer: Teacher trajectory final answer after repair/normalization where applicable.
lexical_class: Lexical outcome label from earlier checks.
repair_version: Conversion/repair version.
turn_count, thinking_count, pull_count, read_count, bash_count, tool_call_count: Tool/process statistics when available.
max_request_context_tokens: Maximum request context length observed during rollout when available.
message_count: Number of messages in the converted training example.
trajectory_path, source_trajectory_path: Original trajectory location when available.
messages: Swift/Qwen3.5 agent SFT message list.

Origin Types

`high_quality_reasoning_core`

Count: 3384 rows.

Source path:

/home/ubuntu/yi/training_survey/sft_dataset_construction/11_sft_training_conversion/outputs/reasoning_core_forced_thinking_qwen35_0706

These are the highest-quality fixed-thinking rows selected for Qwen3.5 thinking-mode SFT. They were exported from the fixed-thinking teacher trajectories and restricted to process_bucket == reasoning_sft_core.

Filtering/cleaning summary:

Started from fixed-thinking GPT-5.4-MINI/NANO-style teacher trajectories generated on the new neutral local-search corpus.
Kept rows judged answer-correct after exact/semantic/repair checks.
Required explicit assistant thinking supervision.
Dropped low/no-thinking auxiliary rows.
Dropped rows with heavy useless terminal behavior such as expensive broad workspace scans.
Dropped low-quality adjacent/repetitive pull patterns.
Required usable final-answer evidence/citation format.
Preserved model-visible reasoning as explicit <think>...</think> blocks.
Converted to Qwen3.5 qwen3_coder agent SFT format with --add_non_thinking_prefix false --preserve_thinking true.

`lower_quality_tool_format_aux`

Count in train-ready view: 1728 rows.

Source path:

/home/ubuntu/yi/training_survey/sft_dataset_construction/11_sft_training_conversion/outputs/fixed_thinking_teacher_clean_v2_llm_numeric_rescue

These rows come from the same fixed-thinking 2026-07-05 teacher trajectory pool but were not part of the strict reasoning core. They are retained as lower-quality/tool-format auxiliary supervision.

Filtering/cleaning summary:

Kept after final-answer correctness and basic process-quality filtering.
Includes low-thinking/tool-format rows that are useful for teaching tool schema, final-answer formatting, and simple local-search behavior.
Passed the teacher_clean_v2 deterministic filters:
- remove heavy bash audit patterns,
- remove poor adjacent sequential pull patterns,
- require standard [@corpus/...] citation format,
- apply numeric-answer consistency checks,
- rescue rows where the LLM numeric judge marked the answer supported.
Final answers had presentation markdown and citation formatting normalized.

`legacy_gpt54_nano_red832_*`

Original source count: 832 rows.

Source path:

/home/ubuntu/yi/swift_sft/runs/qwen35_9b_lora_gpt54_red832_conf70/data

These are the older GPT-5.4-NANO RedSearcher trajectories used for earlier LoRA experiments.

Two versions are represented:

legacy_gpt54_nano_red832_dedup_repaired
- 293 rows in training_dedup_clean_5405.parquet.
- These are the non-overlapping legacy rows retained after de-duplication against fixed-thinking rows and after conversion/repair.
legacy_gpt54_nano_red832_original
- 832 rows in consolidated_three_sources_full_5944.parquet.
- This preserves the full original legacy source for audit/provenance.

The original 832 rows are not all present in the recommended train-ready view because most overlap newer fixed-thinking examples by query_id. The earlier merged export reported:

total rows: 5484
fixed-thinking selected rows: 5191
original legacy rows: 832
legacy rows added after de-duplication: 293
legacy rows skipped due to overlap: 539

Recommended Usage

For training:

from datasets import load_dataset

ds = load_dataset(
    "parquet",
    data_files="training_dedup_clean_5405.parquet",
    split="train",
)

For source analysis:

from datasets import load_dataset

ds = load_dataset(
    "parquet",
    data_files="consolidated_three_sources_full_5944.parquet",
    split="train",
)

Filter by origin:

core = ds.filter(lambda x: x["origin_type"] == "high_quality_reasoning_core")

Rebuild

Run inside the Swift trainer Docker image:

docker run --rm \
  -v /home/ubuntu/yi:/home/ubuntu/yi \
  -w /home/ubuntu/yi/dci_project \
  --entrypoint python \
  swift-sft-qwen35-official:latest \
  /home/ubuntu/yi/training_survey/sft_dataset_construction/12_consolidated_qwen35_sft_sources/src/build_consolidated_dataset.py

Downloads last month: 5

Total file size:

1.21 GB