wchen22/isb1-cc-traces

ISB-1 KV-Cache-Tester Traces

Overview

This dataset is a Hugging Face mirror of datasets/isb1/converted/ from SemiAnalysisAI/InferenceX. It packages the pre-converted ISB-1 multi-turn trace corpus in the JSON schema already consumed by the kv-cache-tester replay path, so downstream users can switch from a local directory to TRACE_DIR=hf_<org>--<repo> with no harness changes.

Current corpus summary:

• 179 validated trace JSON files
• 1,226 total requests across the corpus
• ~9.26 MiB on disk including manifest.json
• Context bands spanning 8k, 32k, 64k, 131k, 500k preview, and 1m preview
• Model coverage including DeepSeek-R1-0528, GPT-OSS-120B, GLM-5, MiniMax-M2.5, and Qwen3.5-397B-A17B variants present in the source corpus

Preferred dataset name:

• semianalysisai/isb1-cc-traces

Fallback if SemiAnalysisAI org write access is unavailable:

• ocwc22/isb1-cc-traces

The intended consumer is Cam's existing kv-cache-tester replay flow, which already accepts hf_<org>--<repo> as a TRACE_DIR source and hydrates the remote dataset locally before replay.

Schema

Each JSON file is one replayable trace object with the following top-level fields:

Field	Type	Notes
id	str	Stable trace/session identifier
models	list[str]	Model family identifiers for the trace
block_size	int	Hash-block size used for hash_ids; current corpus uses 64
hash_id_scope	str	local or global; validated by the stdlib checker
requests	list[object]	Ordered replay requests
tool_tokens	int	Optional token accounting
system_tokens	int	Optional token accounting
totals / isb1	object	Source metadata carried through from conversion

Each request entry includes the fields kv-cache-tester expects today:

• type
• t
• in
• out
• hash_ids
• optional model, stop, input_types, output_types, api_time, think_time

Before publishing or consuming the corpus, validate it with the bundled stdlib-only checker:

python3 tools/validate_kvcache_tester_trace.py datasets/isb1/converted/

Exit codes:

• 0 = all trace files valid
• 1 = one or more trace files failed validation
• 2 = usage or path error

How to use

For zero-friction consumption through the existing replay wrapper, point TRACE_DIR at the HF dataset name in hf_<org>--<repo> form:

TRACE_DIR=hf_semianalysisai--isb1-cc-traces \
bash experimental/multiturn/benchmarks/single_node/multiturn_fp8_h200_trace_replay.sh

That is the whole integration contract: the wrapper handles the HF download and passes the hydrated local mirror into the replay runner. If the preferred org is unavailable, the fallback form is identical:

TRACE_DIR=hf_ocwc22--isb1-cc-traces \
bash experimental/multiturn/benchmarks/single_node/multiturn_fp8_h200_trace_replay.sh

For local inspection after download, the dataset also includes:

• manifest.json — corpus summary and per-trace metadata
• the original directory layout under core/, extension_32k/, extension_64k/, extension_131k/, and preview/

License

This corpus follows the repository license in SemiAnalysisAI/InferenceX, which is Apache-2.0.

Citation

If you use this corpus in a benchmark, report, or derivative evaluation, cite both the InferenceX repository and the dataset name/revision that you consumed. A lightweight citation template is below.

@misc{isb1_kvcache_tester_traces_2026,
  title        = {ISB-1 KV-Cache-Tester Traces},
  author       = {SemiAnalysisAI},
  year         = {2026},
  howpublished = {Hugging Face dataset repository},
  note         = {Preferred repo: semianalysisai/isb1-cc-traces; fallback: ocwc22/isb1-cc-traces}
}