Repository navigation

feat: sync score averaging, adapter scale and recent cleanup - #1

Merged

Abhinavexist merged 39 commits into

mainfrom

sync/adapter-scale-and-docstring-pass

Oct 15, 2026

Merged

feat: sync score averaging, adapter scale and recent cleanup#1
Abhinavexist merged 39 commits into
mainfrom
sync/adapter-scale-and-docstring-pass

Commits on Oct 15, 2026
- docs: ADR-029 plan, what the Score order measurements look at
  
  authored and
  Abhinavexist
  committed
  4edada8
- Measure Score's position effect with lev presentation-checks, and offer Score order averaging (off by default)
  
  authored and
  Abhinavexist
  committed
  eea1476
- ADR-029 evidence: before/after on bench_en, bench_ja as reference; committed presentation report
  
  authored and
  Abhinavexist
  committed
  582c93c
- presentation-checks --compare refuses a report measured under other settings (exit 2)
  
  authored and
  Abhinavexist
  committed
  b4572f4
- Address review: one-read identical control, engine-counted rows and a request row limit
  
  authored and
  Abhinavexist
  committed
  4c27d72
- ADR-029 evidence after review: one-read bias, labelled and calibration effects, row limit
  
  authored and
  Abhinavexist
  committed
  0e5fc6e
- Row limit: refuse only the rows Score averaging adds, never a request main accepts
  
  authored and
  Abhinavexist
  committed
  ff8fbd7
- ADR-029: Score's first-slot rate is a reference, not evidence for averaging
  
  authored and
  Abhinavexist
  committed
  3136a79
- Remove the _eager_model reference that #4's description pointed out
  
  authored and
  Abhinavexist
  committed
  79d9451
- Bound Score-averaged requests by the batcher's token budget, not by rows
  
  authored and
  Abhinavexist
  committed
  b618529
- perf(model): merge the lora adapter when loading
  
  Abhinavexists
  authored and
  Abhinavexist
  committed
  771db93
- feat(modal): sweep the adapter scale on the rating subsets
  
  Abhinavexists
  authored and
  Abhinavexist
  committed
  f1752fb
- docs(decisions): correct the torch.compile rejection in ADR-023
  
  Abhinavexists
  authored and
  Abhinavexist
  committed
  a7e045e
- refactor: tighten comments, docstrings and naming across the repo
  
  Abhinavexists
  authored and
  Abhinavexist
  committed
  724f6e7
- refactor(train): share the checkpoint load and scoring forward
  
  Abhinavexists
  authored and
  Abhinavexist
  committed
  f7f851d
- fix(modal): score the ordinal sweep with levbench's label rule
  
  Abhinavexists
  authored and
  Abhinavexist
  committed
  feb9865
- feat(serve): apply the release adapter scale when loading
  
  Abhinavexists
  authored and
  Abhinavexist
  committed
  281704b
- test(serve): pin the adapter-scale multiplier through the merge
  
  Abhinavexists
  authored and
  Abhinavexist
  committed
  0a2a650
- feat(modal): refit the temperature profile at a chosen adapter scale
  
  Abhinavexists
  authored and
  Abhinavexist
  committed
  e4351bd
- docs(findings): record the adapter-scale result and its selection
  
  Abhinavexists
  authored and
  Abhinavexist
  committed
  ae21270
- refactor(train): derive the bucket window from the training config
  
  Abhinavexists
  authored and
  Abhinavexist
  committed
  5117993
- test(train): cover the ragged final batch in the scored-rows round trip
  
  Abhinavexists
  authored and
  Abhinavexist
  committed
  06b43a8
- refactor(train): tighten the training loop's types
  
  Abhinavexists
  authored and
  Abhinavexist
  committed
  0bce0d3
- test(train): tighten the training-loop tests' types
  
  Abhinavexists
  authored and
  Abhinavexist
  committed
  6efa5d7
- test(packaging): drop the dead help-comment strip and share the doc scan
  
  Abhinavexists
  authored and
  Abhinavexist
  committed
  594e284
- docs(claude): give agents the project context
  
  Abhinavexists
  authored and
  Abhinavexist
  committed
  3031d43
- refactor: replace hand-counted padding with format width specifiers
  
  Abhinavexists
  authored and
  Abhinavexist
  committed
  b8f8760
- fix(levbench): align the jev-vs-llm comparison columns
  
  Abhinavexists
  authored and
  Abhinavexist
  committed
  7dfdd05
- fix(levbench): take the jev model from the sdk default
  
  Abhinavexists
  authored and
  Abhinavexist
  committed
  a107f63
- docs: move internal docstrings to why-comments, keep public api docs
  
  Abhinavexists
  authored and
  Abhinavexist
  committed
  32bb003
- docs(claude): adapt repo references for the public repository
  Abhinavexist
  committed
  15bc6ff
- feat(modal): serve a packaged release with LEV_SERVE_RELEASE
  
  Abhinavexists
  authored and
  Abhinavexist
  committed
  7220285
- docs(setup): document serving a packaged release
  
  Abhinavexists
  authored and
  Abhinavexist
  committed
  7d0fcda
- docs(readme): note score order averaging and the presentation checks
  
  Abhinavexists
  authored and
  Abhinavexist
  committed
  0ff1e83
- feat(release): read the hub write key from .env when publishing
  
  Abhinavexists
  authored and
  Abhinavexist
  committed
  25590d4
- test(release): pin the hub key precedence for publish
  
  Abhinavexists
  authored and
  Abhinavexist
  committed
  b24de30
- docs(readme): report the adapter-scaled release on s1bench
  
  Abhinavexists
  authored and
  Abhinavexist
  committed
  136725c
- docs(findings): record the served adapter-scale run
  
  Abhinavexists
  authored and
  Abhinavexist
  committed
  e4cbb6d
- fix(release): read only HF_TOKEN from ./.env when publishing
  
  Abhinavexists
  authored and
  Abhinavexist
  committed
  e1133c9

Commits

Commits on Oct 15, 2026