reflectio/swe-smith-frozen-trajectories

SWE-Smith Frozen Trajectories

This dataset is a serving-performance workload derived from the tool split of SWE-bench/SWE-smith-trajectories. It is designed for measuring throughput, request rate, time to first token, inter-token latency, and prefix-cache behavior with realistic multi-turn coding agent histories.

It is not a coding-correctness benchmark. The tested model's responses are not executed or scored.

Processing

• Keep trajectories generated by claude-3-7-sonnet-20250219.
• Deterministically sample 300 trajectories with seed 42.
• Turn each source assistant turn into one independent OpenAI-compatible chat request containing the complete original history before that turn.
• Never insert output generated by the model under test into a later request.
• Set each request's max_tokens to the corresponding source assistant output length, capped at 4,096 tokens.
• Stop a trajectory before a request would exceed the 65,536-token context budget.
• Write requests in round-robin trajectory order, preserving turn order within each trajectory while spacing requests from the same trajectory.

Tool-call argument strings were normalized to mappings for the Qwen chat template used during conversion. The prompt/output token counts and context limits are therefore Qwen-oriented and may need to be recomputed for models with materially different tokenizers or chat templates.

Dataset summary

Field	Value
Source trajectories available after model filter	17,715
Sampled / usable trajectories	300 / 300
Requests	10,249
Trajectories truncated at context limit	22
Prompt tokens, P50 / P95 / max	18,346 / 49,481 / 65,352
Effective output tokens, P50 / P95 / max	75 / 854 / 4,096
Total prompt tokens	220,385,156
Total effective output tokens	2,130,488
Exact-LCP cacheable prompt-token share (theoretical)	95.65%

The cacheable share is computed as the exact token-level longest common prefix between consecutive prompts from the same source trajectory. It describes workload reuse potential, not a guaranteed server-side cache-hit rate.

Files

• data/swe_smith_sglang_64k_300sessions.jsonl: one request per line, with messages, max_tokens, and temperature.
• metadata/swe_smith_sglang_64k_300sessions.metadata.jsonl: line-aligned trajectory, turn, token-length, and cache-prefix metadata.
• metadata/swe_smith_sglang_64k_300sessions.stats.json: aggregate statistics and conversion parameters.
• scripts/swe_smith_frozen_converter.py: converter used to create this release.
• SHA256SUMS: file checksums for integrity verification.

The two JSONL files are aligned by request_index: line N in the request file matches line N in the metadata file.

Usage

Load the two configurations independently:

from datasets import load_dataset

requests = load_dataset(
    "reflectio/swe-smith-frozen-trajectories",
    "requests",
    split="train",
    streaming=True,
)
metadata = load_dataset(
    "reflectio/swe-smith-frozen-trajectories",
    "metadata",
    split="train",
    streaming=True,
)

for request, meta in zip(requests, metadata):
    assert meta["request_index"] >= 0
    # Send request directly to an OpenAI-compatible chat-completions endpoint.

For benchmark clients that consume JSONL directly, download the request and metadata files without passing them through the datasets library.

Interpretation and limitations

• Requests contain frozen, semantically coherent source histories; they do not simulate an interactive agent reacting to newly generated model output.
• A client may issue all rows as independent requests, or use trajectory_id to serialize turns per trajectory while running different trajectories in parallel.
• Prefix-cache reuse depends on request routing, scheduling, cache capacity, eviction, and the serving engine's cache implementation.
• The source trajectories can contain unsuccessful attempts; resolved is retained in the metadata.

License and attribution

This derivative dataset follows the upstream dataset's MIT license. Please also review and cite the original SWE-Smith trajectories dataset and its associated project when using this workload.