Skip to main content

Provenance

QuarkLM records corpus provenance in each self-improvement report.

corpus_snapshot.json captures:

  • ledger source ids
  • file paths
  • training permissions
  • curriculum-generation permissions
  • file hashes
  • JSONL record counts
  • admitted memory ids

corpus_diff.json compares the current snapshot to the previous promoted run. This makes corpus changes visible next to weight changes and eval changes.

v0.18 added admission-paraphrase-probes-v0 as an eval-only source.

v0.19 changed the glossary, admitted-memory log, and generated admission probes when learned-ivy-stone entered the corpus.

v0.20 added glossary-probes-v0 as an eval-only source and changed the glossary source to include explicit glossary probe words.

v0.21 added admitted memories learned-noah-shell, learned-ava-coin, and learned-omar-drum; changed admission, admission-paraphrase, glossary, and glossary-probe sources; and kept grammar unchanged.

v0.22 changed grammar, self-probe, and learning-probe sources to teach self-diagnosis source, external-model-shaping policy, and report-evidence repair selection. Admissions and generated admission/glossary probes stayed unchanged.

v0.23 added attempt archives under attempts/attempt-###/, preserving a failed undertrained attempt and a repaired passing attempt in the same run directory.

v0.24 kept corpus sources unchanged from v0.23, passed the archived self-improvement run, and added separate transformer architecture evidence under runs/transformer-v0.24/.

v0.25 added admitted memories learned-mia-ring, learned-leo-kite, and learned-nina-bell; added glossary probe words ring, kite, and bell; changed admission, admission-paraphrase, glossary, and glossary-probe sources; and added transformer architecture evidence under runs/transformer-v0.25/.

v0.26 kept corpus sources unchanged from v0.25, passed the archived self-improvement run, and added transformer answer-lesson evidence under runs/transformer-answer-v0.26/.

v0.27 kept corpus sources unchanged from v0.26, passed the archived self-improvement run, and added context-48 transformer answer evidence plus a faster eval-scoped candidate evaluator under runs/transformer-answer-v0.27/.

v0.28 added admitted memories learned-sara-marble, learned-milo-spoon, and learned-ruth-ribbon; added glossary probe words marble, spoon, and ribbon; changed admission, admission-paraphrase, glossary, and glossary-probe sources; preserved two failed self-improvement attempts before passing attempt-003; and added transformer prefix-choice evidence under runs/transformer-answer-v0.28-choice-prefix-pilot/.

v0.29 kept corpus sources unchanged from v0.28, passed the archived self-improvement run on attempt-001, and added transformer answer-selector evidence under runs/transformer-answer-v0.29-selector-fast/.

v0.30 kept corpus sources unchanged from v0.29, passed the archived self-improvement run on attempt-001, and added transformer selector-emission evidence under runs/transformer-answer-v0.30-selector-emission/.

v0.31 kept corpus sources unchanged from v0.30, passed the archived self-improvement run on attempt-001, and added no-candidate transformer-guided answer-generator evidence under runs/transformer-answer-v0.31-generator-weighted-lr035-80k/.

v0.32 kept corpus sources unchanged from v0.31, passed the archived self-improvement run on attempt-001, and added direct greedy transformer answer-training evidence under runs/transformer-answer-v0.32-direct-base-context32/. Direct transformer loss improved, but strict raw greedy exact answers remained 0/219.

v0.33 kept corpus sources unchanged from v0.32, passed the archived self-improvement run on attempt-001, and added first-error unlikelihood transformer evidence under runs/transformer-answer-v0.33-unlikelihood-context32/. Transformer-only candidate accuracy improved to 37/219, but strict raw greedy exact answers remained 0/219.

v0.34 kept corpus sources unchanged from v0.33, passed the archived self-improvement run on attempt-001, and added staged rollout unlikelihood transformer evidence under runs/transformer-answer-v0.34-staged-context32/. Direct target loss improved and the repeated-output failure changed shape, but strict raw greedy exact answers remained 0/219.

v0.35 kept corpus sources unchanged from v0.34, passed the archived self-improvement run on attempt-001, and added periodic rollout unlikelihood transformer evidence under runs/transformer-answer-v0.35-periodic10-context32/. Candidate discrimination returned to 37/219, but strict raw greedy exact answers remained 0/219.

v0.36 kept corpus sources unchanged from v0.35, passed the archived self-improvement run on attempt-001, and added periodic early-stop unlikelihood transformer evidence under runs/transformer-answer-v0.36-periodic-earlystop10-context32/. Candidate discrimination stayed at 37/219, answer-target NLL improved to 2.9311, and strict raw greedy exact answers remained 0/219 with a repeated " a" loop.

v0.37 kept corpus sources unchanged from v0.36, passed the archived self-improvement run on attempt-001, and added periodic repeat-loop unlikelihood transformer evidence under runs/transformer-answer-v0.37-periodic-repeat50-context32/. Candidate discrimination stayed at 37/219, answer-target NLL improved to 2.9041, and strict raw greedy exact answers remained 0/219 with a repeated " t" loop.

v0.38 kept corpus sources unchanged from v0.37, passed the archived self-improvement run on attempt-001, and added periodic balanced repair transformer evidence under runs/transformer-answer-v0.38-periodic-balanced50-context32/. Candidate discrimination stayed at 37/219, answer-target NLL improved to 2.8552, and strict raw greedy exact answers remained 0/219 with a repeated " t" loop.

v0.39 kept corpus sources unchanged from v0.38, passed the archived self-improvement run on attempt-001, rejected two generated-prefix recovery pilots, and added periodic sequence-repair transformer evidence under runs/transformer-answer-v0.39-periodic-sequence50-context32/. Candidate discrimination stayed at 37/219, answer-target NLL improved to 2.8257, and strict raw greedy exact answers remained 0/219 with a repeated " t" loop.

v0.40 kept corpus sources unchanged from v0.39, passed the archived self-improvement run on attempt-001, rejected a loop-escape-only pilot, kept a protected sequence-plus-loop pilot as non-selected evidence, and added branch repair transformer evidence under runs/transformer-answer-v0.40-branch-context32/. Candidate discrimination stayed at 37/219, answer-target NLL improved to 2.5427, and strict raw greedy exact answers remained 0/219 with a repeated "ten" loop.

v0.41 kept corpus sources unchanged from v0.40, passed the archived self-improvement run on attempt-001, rejected a full-dose branch-contrast pilot, and added sparse branch-repair/contrast transformer evidence under runs/transformer-answer-v0.41-branch-repair-contrast50-context32/. Candidate discrimination stayed at 37/219, answer-target NLL improved to 2.4734, and strict raw greedy exact answers remained 0/219 with a repeated "te"/"e" loop.

v0.42 kept corpus sources unchanged from v0.41, passed the archived self-improvement run on attempt-001, and widened the sparse branch-contrast transformer evidence under runs/transformer-answer-v0.42-branch-repair-contrast50-dim8-context32/. Candidate discrimination stayed at 37/219, answer-target NLL improved to 2.4129, and strict raw greedy exact answers remained 0/219 with the short wrong completion " te.".

Post-v0.42 unpromoted transformer work kept corpus sources unchanged and added runtime, diagnosis, and rejected checkpoint evidence. The transformer forward pass now computes only the final position used by the language-model head. Answer-training artifacts now record prompt context-coverage metrics. The hard-negative context-32 run runs/transformer-answer-v0.43-hard-branch-contrast4-dim8-context32/ preserved 37/219 candidate discrimination but regressed loss, NLL, and greedy output. The context-80 run runs/transformer-answer-v0.43-branch-repair-contrast50-dim8-context80/ covered all semantic eval templates (219/219) but still trailed v0.42 on direct loss and answer NLL. The 1500-step context-80 run reached 38/219 candidates but regressed other promotion metrics. Optional layer normalization was added as a tested architecture flag, but the context-80 screen runs/transformer-answer-v0.43-layernorm-screen-dim8-context80/ preserved only 37/219 candidates and regressed answer NLL with repeated " y"/"e" greedy loops. Branch-span repair was added as a tested direct-answer policy, but runs/transformer-answer-v0.43-branch-span3-screen-dim8-context32/ preserved only 37/219 candidates and regressed answer NLL with a long "neeee" greedy loop. Multi-layer transformer support was added as a tested architecture option, but runs/transformer-answer-v0.43-two-layer-screen-dim8-context32/ was interrupted before final direct-answer metrics because the full-block scalar autograd path was too slow for the regular loop. The final stacked layer was then optimized to compute only the last state with logit-equivalence coverage, but runs/transformer-answer-v0.43-two-layer-finalopt-screen-dim8-context32/ was still interrupted before final metrics because intermediate full-state training remains too expensive. A later bounded screen added top-layer-only direct-answer updates for stacked transformers plus the explicit --skip-post-direct-snapshot control: runs/transformer-answer-v0.43-two-layer-toponly-skip-screen-dim8-context32/ completed, saved a checkpoint, recorded that the post-direct candidate snapshot was skipped, improved direct-answer target loss from 3.5186 to 3.2436, and still failed direct greedy exact at 0/219 with repeated "a" output. It is runtime and training-loop evidence, not promotion evidence. Direct-answer snapshots then gained branch-profile diagnostics under runs/transformer-answer-v0.43-branch-profile-smoke-dim4-context16/, recording the model's own branch-position logits, dominant predicted tokens, target-token distribution, and target margin. The smoke profile showed QA branch accuracy staying at 1/8 while dominant prediction moved from all "o" to all "y", which is prompt-independent branch-collapse evidence. Branch-collapse repair then used the dominant sampled branch token as the unlikelihood negative. The full-dose smoke at runs/transformer-answer-v0.43-branch-collapse-smoke-dim4-context16/ regressed loss and moved collapse to all "a" predictions. The periodic smoke at runs/transformer-answer-v0.43-periodic-branch-collapse-smoke-dim4-context16/ improved direct loss to 3.5157, but branch accuracy stayed 1/8 and the dominant prediction moved to all "n". Branch-batch contrast then trained several distinct target branches in one update. The full-dose smoke at runs/transformer-answer-v0.43-branch-batch-smoke-dim4-context16/ improved loss only slightly and moved collapse to all "y" predictions. The periodic smoke at runs/transformer-answer-v0.43-periodic-branch-batch-smoke-dim4-context16/ improved direct loss to 3.5248, but QA branch accuracy regressed to 0/8 and the dominant prediction moved to all "a". A representation-side context-mean option was then added without changing corpus sources. The branch-batch screen runs/transformer-answer-v0.43-context-mean-branch-batch-smoke-dim4-context16/ improved direct loss to 3.5252, and the branch-repair screen runs/transformer-answer-v0.43-context-mean-branch-repair-smoke-dim4-context16/ improved direct loss to 3.5310; both regressed QA branch accuracy to 0/8 and collapsed the dominant prediction to all "a". A learned context-projection option followed, again without changing corpus sources. The branch-repair screen runs/transformer-answer-v0.43-context-projection-branch-repair-smoke-dim4-context16/ improved direct loss to 3.5217, and the branch-batch screen runs/transformer-answer-v0.43-context-projection-branch-batch-smoke-dim4-context16/ improved direct loss to 3.5252; both moved their projection weights, regressed QA branch accuracy to 0/8, and collapsed the dominant prediction to all "a". A prompt-attention summary option followed, again without changing corpus sources. The branch-repair screen runs/transformer-answer-v0.43-prompt-attention-branch-repair-smoke-dim4-context16/ improved direct loss to 3.5217, and the branch-batch screen runs/transformer-answer-v0.43-prompt-attention-branch-batch-smoke-dim4-context16/ improved direct loss to 3.5252; both moved their zero-initialized output projection weights, regressed QA branch accuracy to 0/8, and collapsed the dominant prediction to all "a". Branch-context coverage diagnostics were then added to direct-answer snapshots without changing corpus sources. The context-16 screen runs/transformer-answer-v0.43-branch-context-coverage-smoke-dim4-context16/ showed QA branch contexts had 0/8 semantic coverage and 4 ambiguous branch windows. The context-32 screen runs/transformer-answer-v0.43-branch-context-coverage-smoke-dim4-context32/ removed QA ambiguity but still had 0/8 semantic coverage. The context-80 screen runs/transformer-answer-v0.43-branch-context-coverage-smoke-dim4-context80/ reached complete branch-context coverage across all eval sets (219/219) with zero ambiguous branch contexts. The branch-context gate was then added as an opt-in training guardrail, again without changing corpus sources. The context-16 gate screen runs/transformer-answer-v0.43-branch-context-gate-smoke-dim4-context16/ failed the required gate and recorded actual_steps: 0 for 5 requested direct-answer steps. The context-80 gate screen runs/transformer-answer-v0.43-branch-context-gate-smoke-dim4-context80/ passed the required gate and recorded actual_steps: 1 for 1 requested direct-answer step. Branch-only direct-answer snapshots were then added as an explicit screening mode for longer-context repair runs, again without changing corpus sources. The context-80 gated branch-only screen runs/transformer-answer-v0.43-branch-context-gated-branchonly-smoke-dim4-context80/ passed the required gate across all 219/219 semantic records, ran all 5 requested direct-answer steps, and recorded skipped greedy evals while retaining branch profiles and branch-context gate evidence. Two dim8 context-80 branch-only follow-up screens then tested the best prior sparse repair/contrast policy and branch-batch contrast under complete branch context. The repair/contrast screen runs/transformer-answer-v0.43-branchonly-periodic-repair-contrast50-dim8-context80/ ran all 100 requested direct steps and lowered interval train loss, but final QA branch prediction collapsed to all "a". The branch-batch screen runs/transformer-answer-v0.43-branchonly-branch-batch-dim8-context80/ ran all 50 requested direct steps and lowered interval train loss further, but also collapsed QA branch prediction to all "a". Branch diversity was then promoted from narrative diagnosis to an explicit snapshot target, again without changing corpus sources. The smoke run runs/transformer-answer-v0.43-branch-diversity-target-smoke-dim4-context80/ passed the branch-context gate, ran all 5 requested direct steps, and recorded branch_diversity_target failure across all 9 multi-target eval profiles. The first diversity-aware training objective was then added as branch-diversity-unlikelihood, still without changing corpus sources. The context-80 smoke runs/transformer-answer-v0.43-branch-diversity-train-smoke-dim4-context80/ passed the branch-context gate and ran all 10 requested direct steps, but the diversity target still failed across all 9 multi-target eval profiles. Output-bias freezing was then added as a direct-answer stabilizer, still without changing corpus sources. The context-80 smoke runs/transformer-answer-v0.43-branch-diversity-freezebias-smoke-dim4-context80/ passed the branch-context gate and ran all 50 requested direct steps with --direct-answer-freeze-output-bias, but the diversity target still failed across all 9 multi-target eval profiles. A restricted branch-target softmax objective followed, still without changing corpus sources. The context-80 smoke runs/transformer-answer-v0.43-branch-target-softmax-freezebias-smoke-dim4-context80/ passed the branch-context gate, froze output bias, and ran all 50 requested direct steps. It briefly raised QA predicted diversity to two tokens at step 20, but the final diversity target still failed across all 9 multi-target eval profiles. Best-branch-snapshot restoration followed, still without changing corpus sources. The context-80 smoke runs/transformer-answer-v0.43-branch-target-softmax-restorebest-smoke-dim4-context80/ passed the branch-context gate, froze output bias, ran all 50 requested direct steps, and restored the final checkpoint from step 40; the final diversity target still failed across all 9 multi-target eval profiles. Prompt-prefix projection followed, still without changing corpus sources. The context-80 smoke runs/transformer-answer-v0.43-prompt-prefix-target-softmax-restorebest-smoke-dim4-context80/ passed the branch-context gate, moved all 20 prompt-prefix projection parameters, and restored the final checkpoint from step 40; the final diversity target still failed across all 9 multi-target eval profiles. Prompt-position projection followed, still without changing corpus sources. The context-80 smoke runs/transformer-answer-v0.43-prompt-position-target-softmax-restorebest-smoke-dim4-context80/ passed the branch-context gate, moved 1108/1284 prompt-position projection parameters, and restored the final checkpoint from step 40; the final diversity target still failed across all 9 multi-target eval profiles. A pairwise branch-target margin objective followed, still without changing corpus sources. The prompt-position context-80 smoke runs/transformer-answer-v0.43-branch-target-margin-prompt-position-smoke-dim4-context80/ passed the branch-context gate, ran all 50 direct steps, moved train loss 4.8973 -> 4.7784, moved 1108/1284 prompt-position projection parameters, and restored the final checkpoint from step 40; the final diversity target still failed across all 9 multi-target eval profiles. Branch representation diagnostics and contrastive hidden-state training followed, still without changing corpus sources. The high-weight prompt-position context-80 smoke runs/transformer-answer-v0.43-branch-representation-contrast50-prompt-position-smoke-dim4-context80/ recorded hidden-state distance profiles, ran all 50 direct steps with --direct-answer-contrast-weight 50.0, and restored the final checkpoint from step 40; the final diversity target still failed across all 9 multi-target eval profiles. A dim-8 capacity screen followed, still without changing corpus sources. The completed 40-step prompt-position context-80 smoke runs/transformer-answer-v0.43-branch-representation-contrast50-prompt-position-smoke-dim8-context80-steps40/ used embedding/feed-forward dimensions 8/16, restored the final checkpoint from step 10, and increased measured QA hidden distance; the final diversity target still failed across all 9 multi-target eval profiles. Prompt-position scale screening followed, still without changing corpus sources. The scale-32 context-80 smoke runs/transformer-answer-v0.43-prompt-position-scale32-repcontrast50-smoke-dim4-context80/ ran all 50 direct steps, moved 1108/1284 prompt-position projection parameters, restored the final checkpoint from step 40, and increased restored QA hidden distance to about 0.01235; the final diversity target still failed across all 9 multi-target eval profiles. The next checkpoint records an engineering-only open-source structure audit in STRUCTURE_AUDIT.md: QuarkLM may study model/trainer/tokenizer/checkpoint patterns, but no external weights, tokenizer vocabularies, embeddings, datasets, or training text enter the corpus or learned artifacts. The audit selects an opt-in pre-layer-norm transformer block path with final normalization as the next structural screen before another branch-loss repair. That path was implemented and screened in runs/transformer-answer-v0.44-prelayernorm-repcontrast50-prompt-position-smoke-dim4-context80/ without changing corpus sources. The run moved prompt-position and final-norm parameters and cracked full collapse in 7/9 multi-target profiles, but QA and heldout remained collapsed and the final diversity target still failed across all 9 multi-target eval profiles. A target-balanced branch-batch screen followed, still without changing corpus sources. The run runs/transformer-answer-v0.44-target-balanced-prelayernorm-repcontrast50-prompt-position-smoke-dim4-context80/ used target-bucket branch batches, ran all 50 direct steps, and restored the final checkpoint to baseline step 0 because trained snapshots scored worse; all 9/9 multi-target eval profiles remained collapsed to one global token. None of these runs were promoted.