Input should contain retrieved paper evidence selected by the AgentKernel paper QA harness. The model is not a broad standalone scientific oracle; it is a retrieved-evidence synthesizer.

Expected answer shape:

Answer: [P1] ... <AK_END>

Audit

Native long-cap generation audit: stage7949_v27_paper_qa_ak_end_native_generation_longcap_audit.

Heldout eval + strict rows: 50.

Key metrics:

starts_answer_rate: 1.0
cites_p1_rate: 1.0
ak_end_seen_rate: 1.0
ends_punctuation_rate: 1.0
target_f1_mean: 0.7065
internal_token_leak_rate: 0.0
quoted_title_like_rate: 0.0
ak_end_prob_at_target_boundary_mean: 0.999986
ak_end_rank_at_target_boundary_median: 1.0

Files

model/config.json
model/model.safetensors
tokenizer/tokenizer.json
tokenizer/tokenizer_config.json
agentkernel_lite_encdec_manifest.json
audits/stage7947_train_summary.json
audits/stage7949_native_generation_audit.json

Notes

The browser BitNet export was not included in this repo because local export failed after dense checkpoint save due a missing runtime file: browser/bitnet/bitnet_webgpu.js. This repo is the audited dense checkpoint package.

PeytonT/agentkernel-100m-paper-qa-synthesizer-stage7947

AgentKernel 100M Paper QA Synthesizer Stage 7947

Intended Use

Audit

Files

Notes

PeytonT/
agentkernel-100m-paper-qa-synthesizer-stage7947