You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

edgebench-bit-exact-source-raw

Raw-char (plaintext, model-agnostic) export of the EdgeBench bit-exact strict-replay corpus: recorded agentic sessions (trie replay schema v2) with roles, message content, reasoning, tool calls/results, timing (wait/replace_context), per-event max_tokens, and capture metadata — and NO model-specific token ids (the GLM-5.2 recorded_prompt_token_ids / recorded_output_token_ids fields are stripped; 36,014 fields removed across 18,007 generate events).

Access is gated: request access with your affiliation and purpose. The data contains recorded agent-session content and is intended for serving-performance research only, not for training.

Lineage

  • Derived read-only from xp-ebexact-glm52-94e587 (GLM-5.2 bit-exact capture; source model zai-org/GLM-5.2-FP8 revision ba978f7d347eaf65d22f1a86833408afdb953541), 137 main traces + 2 smoke traces, 18,007 requests across smoke + wave-1..3.
  • Exporter: build_raw_char_export.py v1.0 — strips only the four token-id fields from generate events and re-serializes rows canonically (json.dumps(..., separators=(",", ":"), sort_keys=True)). Everything character-level is preserved verbatim.

Round-trip contract

This export is the character-level SOURCE for token-authoritative renders. Running retokenize_traces_dsv4.py v1.1 (published in weili-0234/edgebench-bit-exact-dsv4 lineage docs) over these waves with the deepseek-ai/DeepSeek-V4-Flash-DSpark snapshot (62af8fffb2f7030cac4de2f0169f5b8d1101b646), its official encoding/encoding_dsv4.py encoder, and RETOKENIZE_AT=2026-08-20T04:00:00+00:00 reproduces the published DSV4 waves byte-identically up to exactly two per-trace provenance strings (metadata.retokenize.tool version and .retokenized_at, which carry the original build's wall-clock timestamps in the published corpus) — both halves of this contract are cmp-verified at publish time; see the DSV4 dataset's build-report.md. The same recipe with a different target tokenizer/template yields a token-authoritative corpus for that target.

NOTE: these rows are NOT directly strict-replayable (no token ids); replay requires a render step (or a text-fallback replayer).

Contents: protected/{smoke,wave-1,wave-2,wave-3}.schema-v2.jsonl, protected/*.manifest.json, source-provenance.json, receipt.sha256.

Downloads last month
18