README.md

---
language:
- en
- zh
license:
- cc-by-4.0
task_categories:
- tabular-regression
tags:
- large-language-models
- distributed-training
- performance-modeling
- profiling
- mindspeed-llm
- ascend
- parallelism
configs:
- config_name: default
  data_files:
  - split: train
    path: qwen3_8b_4k/step_time_summary.csv
---

# LM-SimBench (Example Snapshot)

## Dataset Description

This repository distributes a **compact example snapshot** of LM-SimBench, the structured CSV release of large-scale LLM training-performance profiling data. The snapshot is provided so reviewers and readers can inspect file layout, schemas, and representative records **without downloading the multi–tens-of-gigabyte full release**.

The profiling methodology, software stack, and field definitions are the same as in the complete LM-SimBench dataset:

**Full release:** [anonymous-neurips26-ljasd/LM-SimBench](https://huggingface.co/datasets/anonymous-neurips26-ljasd/LM-SimBench)

Organized CSV fields correspond to the dataset card of that repository. The original raw traces are indexed from the companion raw-data repository linked there (see **Reference(s)** on the full dataset card).

This snapshot is intended for **research inspection and format verification**. It is **not** a downsized substitute for experiments that require full corpus coverage or model diversity. For modeling and benchmarking at scale, use the full LM-SimBench release above.

The snapshot is released under the same license as the full dataset (see **License/Terms of Use**).

## Relationship to the Full LM-SimBench Release

| Aspect | This repository (`LM-SimBench_example`) | Full release (`LM-SimBench`) |
| --- | --- | --- |
| Scope | Single model directory (`qwen3_8b_4k`), fixed subset of feasible parallel configurations | Multiple model families and contexts; full feasible-configuration coverage per model |
| Repository-level `data_manifest.csv` | Not included (not required for a path-preview snapshot) | Included for corpus-wide indexing |
| CSV schemas (overlapping files) | Same column semantics as documented on the full dataset card | Canonical documentation |

## Contents of This Snapshot

**Model directory:** `qwen3_8b_4k`

**Model-level CSV files** (same roles as in the full release where names overlap):

- `step_time_summary.csv` — configuration-level iteration time.
- `feasible_configs_sorted.csv` — feasible configurations sorted by `iteration_time_ms` within this model directory.
- `peak_memory.csv` — rank-level peak allocated memory.
- `operator_latency_by_category.csv` — operator latency by category and type.

**Additional model-level CSV files in this snapshot** (derived from the same preprocessing pipeline; useful for debugging and fine-grained analysis; **not** listed on the main LM-SimBench dataset card):

- `step_trace_time_summary.csv` — rank-level breakdown of iteration time into coarse phases (e.g., compute, communication, overlap-related fields as emitted by the profiler pipeline).
- `comm_summary.csv` — rank- and configuration-level **aggregated** communication statistics (sums/means of elapsed and wait-related quantities, link-type aggregates, and related summaries), summarizing content related to per-event `comm_time.csv` files below.

**Per-configuration layout** (unchanged from the full release):

```text
qwen3_8b_4k/
  step_time_summary.csv
  feasible_configs_sorted.csv
  peak_memory.csv
  operator_latency_by_category.csv
  comm_summary.csv
  step_trace_time_summary.csv
  <parallel_config>/
    rank_<id>/
      comm_time.csv
      matrix_operator_latency.csv
```

**Parallel configurations included:** this snapshot contains **26** configuration directories under `qwen3_8b_4k`, following the naming convention `dp<dp>-tp<tp>-pp<pp>-cp<cp>-<RC|noRC>` documented on the full dataset card.

**Approximate total size:** on the order of **2 GB** (exact byte size depends on Hugging Face packaging).

## Dataset Fields (Overlapping Files)

Field definitions for `step_time_summary.csv`, `feasible_configs_sorted.csv`, `peak_memory.csv`, `operator_latency_by_category.csv`, `comm_time.csv`, and `matrix_operator_latency.csv` are **identical** to those in the full LM-SimBench dataset card:

https://huggingface.co/datasets/anonymous-neurips26-ljasd/LM-SimBench

Readers should consult that card for authoritative column-level documentation.

### step_trace_time_summary.csv (this snapshot only)

Rank-level summary of iteration-time attribution (microseconds in the released files), including columns such as computing, communication-related buckets, pipeline stage and bubble-related fields, and other profiler-phase labels as produced by the extraction tooling. Exact column names appear in the file header.

### comm_summary.csv (this snapshot only)

Rank-level aggregates of communication behavior per configuration, including elapsed-time summaries, wait- and synchronization-related aggregates where present, per-link-type transit summaries, and matrix-level communication summaries where applicable. Exact column names appear in the file header.

## Intended Usage

- Verify CSV schemas and directory layout before retrieving the full LM-SimBench release.
- Lightweight reproduction of loading and parsing code in restricted download environments.
- Teaching and onboarding for the profiling data format.

For substantial empirical studies, use the **full** LM-SimBench dataset linked above.

## License/Terms of Use

This snapshot is governed by the **Creative Commons Attribution 4.0 International License (CC BY 4.0)**, consistent with the full LM-SimBench release.

Users may share and adapt the data subject to attribution and license compatibility requirements stated on the full dataset card.

## Limitations

- **Subset only:** statistics and model coverage do not match the full corpus.
- **Environment specificity:** measurements reflect the stack and hardware configuration described on the full dataset card.
- **Profiler artifacts:** data may include measurement overhead and runtime variability as in the full release.
- **Extra CSV files** (`step_trace_time_summary.csv`, `comm_summary.csv`) are ancillary to the main card documentation; downstream tools should not assume their presence in every third-party mirror of LM-SimBench unless explicitly stated.

## Dataset Characterization

**Data collection method:** Automated profiler collection from distributed LLM training runs (same pipeline as the full release).

**Labeling method:** Automated extraction from training logs and profiler records.

## Ethical Considerations

As with LM-SimBench, this snapshot contains **performance profiling metrics**, not natural-language conversations, personal identifiers, or user-generated text. Downstream use must comply with the dataset license, hosting platform terms, and applicable policies.