Jazhyc/Gleipnir-Prefix-Teacher-Cache

Gleipnir Prefix Teacher Cache

Research artifact containing 133,947 intermediate tool-trajectory predictions from Qwen/Qwen3.5-27B-FP8. This is a prediction cache, not a self-contained trajectory dataset. It contains no trajectory text, original hard labels, privileged rationales, or Kimi K3 full-trajectory targets.

Important numerical limitation

The completed cache failed its numerical-agreement audit. On a fixed 64-prefix sample, fresh versus cached probability differences had mean absolute error 0.010255635 and maximum 0.120145498; five exceeded the predeclared 0.05 maximum limit. An earlier 24-prefix startup test had passed (mean 0.003717, maximum 0.030967). Do not interpret that small passing test as validating this full cache. The original failed audit and a five-case replay diagnostic are included unchanged. Repeated fresh predictions matched exactly on those five cases, while growing-prefix replay remained sensitive to execution history.

The project owner explicitly authorized an exploratory training screen using these imperfect targets without recaching. That authorization is not evidence that the numerical issue was fixed, or that the targets are calibrated. No training benefit is established by this release. Not a production safety system.

Contents and provenance

• logits.jsonl: one JSON object per intermediate boundary.
• contract.json: frozen model revision, runtime, prompt, source and reference hashes.
• complete.json: completed coverage and annotation runtime.
• teacher_prefix.txt: exact teacher instruction, evaluating behavior visible so far.
• fresh_audit.json, failure_replay.json: numerical reliability evidence.
• SHA256SUMS.json: checksums and sizes of release files.

The teacher revision is 97f5941bf617e31c5e237364a8602ce3f03a551a. Annotation used two H100 SXM5 GPUs, tensor parallelism 2, upstream FP8 weights, BF16 model dtype, vLLM 0.28.0, Torch 2.13.0, Transformers 5.14.1, FlashInfer GDN prefill, aligned prefix caching, and one active growing trajectory. Runtime recurrent-state cache dtype was FP32. Responses were constrained to one decision token (0 or 1), with both token logprobs explicitly requested.

Annotation completed on 2026-09-06 in 27,508 seconds: 1,199,840,658 input tokens, of which 1,065,849,568 were reported cached (88.83%). Source counts: STRIDE 26,791; BashArena 72,894; BashBench 27,440; Gloom 6,822. Final full-trajectory boundaries and statement-only turns were excluded; those statements remain visible in later context. Twelve malformed boundaries were not guessed. Format warnings are retained, including compressed step sequences.

Schema and use

logprob_0 and logprob_1 are raw returned decision logprobs. score is the binary-renormalized probability of problematic behavior: exp(logprob_1) / (exp(logprob_0) + exp(logprob_1)) (use a stable softmax). These scores are teacher judgments, not ground truth.

parent_prompt_id and lineage_group identify the local parent/lineage; source identifies the source family. These IDs are not universal upstream row IDs. prefix_ordinal is zero-based; prefix_count counts eligible prefixes. end_character is an exclusive Python-string character offset, not a byte or token offset. trajectory_sha256 and prefix_sha256 identify exact UTF-8 text. Other fields record rendered prompt hashes, contract identity, timestamps, prompt/completion/cache token counts, and parser warnings.

To join against an independently obtained matching parent artifact:

1. Verify the parent-file SHA256 against contract.json before trusting local IDs.
2. Extract the text after <agent_trajectory>\n and before the final </agent_trajectory> in student_prompt, without stripping whitespace.
3. Verify trajectory_sha256, slice trajectory[:end_character], then verify prefix_sha256. SHA256 inputs are UTF-8 encoded strings.
4. Render teacher_prefix.txt.rstrip() + "\n<agent_trajectory>\n" + prefix, add a newline if the prefix lacks one, then append </agent_trajectory>\n. Verify rendered_user_prompt_sha256.
5. Apply the pinned Qwen tokenizer's user chat template with add_generation_prompt=True, enable_thinking=False, append Prediction:, and verify serving_prompt_sha256.

The parent source is not distributed here; users without the matching source cannot reconstruct trajectory text from these hashes. Preserve grouped splits and avoid treating overlapping prefixes as independent evaluation examples.

Rights and access

This release distributes derived numeric predictions and alignment metadata, not the underlying source corpora. It does not grant rights to source datasets, remove access restrictions, or relicense upstream material. No blanket license for the underlying trajectories is asserted. Obtain source data separately and follow its applicable terms. The original Kimi targets are not part of this release.