RomanNum3ral
/
soup
mirror of https://github.com/razor-ai/soup.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Commit Graph

  • 9117da1054 docs: soup train has no --task flag main Alpamys 2026-08-05 17:29:24 +0500
  • 8f9c8435b1 chore: apply pre-commit whitespace fixes Alpamys 2026-08-05 02:48:49 +0500
  • 77eb917ce5 fix(data): make `soup data demo` metadata match the fixtures it ships Alpamys 2026-08-05 02:47:22 +0500
  • 8c2bad55a3 test(examples): parse every config in examples/configs/ Alpamys 2026-08-05 02:36:35 +0500
  • d75797cc3e docs(examples): make every command and promise in examples/ true Alpamys 2026-08-05 02:36:19 +0500
  • 66be481ade fix(examples): migrate example configs to the current schema Alpamys 2026-08-05 02:35:53 +0500
  • d0b7c623ff ci: enable the pre-commit hooks, minus ruff-format Alpamys 2026-08-05 02:34:33 +0500
  • 1b85c2bc51 ci: one retry budget for the whole warm list, and let any cell save Alpamys 2026-08-05 02:05:32 +0500
  • 64a86bbc5a ci: lint the workflows, so an invalid one fails loudly instead of silently Alpamys 2026-08-05 01:37:40 +0500
  • 7d3be55d66 ci: HF_HOME cannot be set from the job-level env block Alpamys 2026-08-05 01:27:27 +0500
  • 195d60b8c7 ci: cache the HF Hub models instead of re-downloading them every run Alpamys 2026-08-05 01:23:09 +0500
  • b059f1ca5d fix(streaming): panel was titled after a flag that does not exist Alpamys 2026-08-05 00:09:56 +0500
  • 115f29df15 docs(readme): show the run, not just the claim Alpamys 2026-08-04 23:50:07 +0500
  • c6ab58ded7 docs(readme): lead with the 8B-on-4GB claim above the fold Alpamys 2026-08-04 23:37:17 +0500
  • 780daa1417 docs: restore the donation link alongside the hardware ask Alpamys 2026-08-04 21:59:23 +0500
  • 6cb84e7973 fix(deps): both trl bounds were wrong — >=0.14.0,<0.27, settled by construction Alpamys 2026-08-03 22:36:39 +0500
  • b62c1dd294 docs(readme): DOI badge was not rendering — serve it from shields.io Alpamys 2026-08-03 21:19:16 +0500
  • 6575355bfd fix(tests): actually apply the two skips — the previous commit applied neither Alpamys 2026-08-03 20:29:14 +0500
  • cb44d9e64e fix(tests): three environment-specific failures in the v0.72.4 suite Alpamys 2026-08-03 20:08:58 +0500
  • 1ca229fdbf fix(deps): trl removed max_prompt_length in stages — cap <0.25, not <0.29 Alpamys 2026-08-03 19:46:27 +0500
  • e58c11b440 fix(deps): the trl break is at 0.29.0, not 1.0 — correct the cap Alpamys 2026-08-03 19:22:23 +0500
  • 35b59346b3 fix(deps): cap trl below 1.0 — six preference trainers cannot build on it Alpamys 2026-08-03 19:00:24 +0500
  • 4fb25404f9 feat(streaming): preference losses over layer streaming (v0.72.4) Alpamys 2026-08-03 18:38:13 +0500
  • 43bd776b5f docs: add the layer-streaming preprint DOI Alpamys 2026-08-03 14:45:54 +0500
  • 2210ca83e3 docs: swap the Discord invite for a non-expiring one Alpamys 2026-08-01 16:58:35 +0500
  • fba981595b docs: add the Discord server and split the project and personal contacts Alpamys 2026-08-01 16:55:24 +0500
  • 2bf714bec8 docs: add a Contact section to the README Alpamys 2026-08-01 14:58:13 +0500
  • dd9818e8f3 docs: use makazanalpamys@gmail.com as the maintainer contact Alpamys 2026-08-01 14:51:23 +0500
  • b060c99a7f docs: drop the Stripe donation link and point support at hardware instead Alpamys 2026-08-01 14:11:08 +0500
  • d73eb1c843 docs(benchmarks): publish the layer-streaming measurement records Alpamys 2026-07-31 15:54:14 +0500
  • 70d1d783a4 fix(tests): NF4 fixtures used a degenerate hidden_size that CPU bitsandbytes rejects Alpamys 2026-07-29 00:10:00 +0500
  • 7b925762f2 fix(deps): cap the [mcp] extra below 2.0 Alpamys 2026-07-28 23:39:01 +0500
  • bcbf72e586 feat(train): layer streaming breadth — 6 more archs, bigger batches, resume, disk tier (v0.72.3) Alpamys 2026-07-28 23:15:20 +0500
  • 1c7540c314 docs: fix two broken anchor links in the layer-streaming pages Alpamys 2026-07-28 16:29:09 +0500
  • 7c398ccdfd test(stream): skip the two NF4 training-step tests when MPS is the accelerator Alpamys 2026-07-28 15:32:56 +0500
  • 08343f8f35 feat(train): NF4 layer streaming — fine-tune Llama-3.1-8B on a 4 GB card (v0.72.2) Alpamys 2026-07-28 15:15:20 +0500
  • de6add9f34 docs: add a donation link and a Sponsor button Alpamys 2026-07-28 12:35:29 +0500
  • 91ec40bd4f test(cli): cover light cores that only a command body imports Alpamys 2026-07-28 12:22:02 +0500
  • 6e79e5e1bf test(cli): extend the startup guard to command modules and real invocations Alpamys 2026-07-28 12:19:39 +0500
  • 4bdd6ee237 perf(cli): stop importing PyTorch on CLI startup (soup --help 6.0s -> 1.15s) Alpamys 2026-07-28 12:07:00 +0500
  • af8683b61d fix(train): streamed adapters were saved unloadable (v0.72.1) Alpamys 2026-07-27 23:18:11 +0500
  • 4dae34a6a4 docs(v0.72.0): list layer streaming in the README docs index and commands reference Alpamys 2026-07-27 01:06:53 +0500
  • 72be45db85 fix(tests): two v0.72.0 tests were platform-dependent, not portable Alpamys 2026-07-27 00:42:42 +0500
  • 00833ac789 feat(train): layer streaming — fine-tune models larger than VRAM (v0.72.0 BETA) Alpamys 2026-07-26 23:58:06 +0500
  • bfa2093ce8 test(online-dpo): smoke-verify train() logs a rewards/* metric (#300) Alpamys 2026-07-19 23:15:14 +0500
  • 481dc388d3 fix(sft): mirror vision-processor pad_token from nested tokenizer (#302) Alpamys 2026-07-19 23:10:35 +0500
  • e9266825e3 feat(adapters): exact concat+SVD task-arithmetic for mixed-rank LoRA (#305) Alpamys 2026-07-19 22:59:19 +0500
  • 10e5522255 fix(callbacks): make ReLoRACallback + HFPushCallback real TrainerCallback subclasses (#308) Alpamys 2026-07-19 22:49:32 +0500
  • 03282e9ef6 feat(reward): soup reward stress — adversarial verifier gameability probe (v0.71.41) Alpamys 2026-07-19 20:53:37 +0500
  • ea3325ea09 docs: sync CONTRIBUTING structure + docs index for v0.71.40 Alpamys 2026-07-19 19:17:16 +0500
  • 43a7fb21a5 feat(reward): soup reward synth — synthesize a deterministic reward verifier (v0.71.40) Alpamys 2026-07-19 18:50:47 +0500
  • cba984a574 feat(ship): close the evidence loop — emit-evidence + config + provenance + PR comment (v0.71.39) Alpamys 2026-07-19 12:30:35 +0500
  • f6bc8e7bdd feat(ship): make soup ship's leg-2 regression gate real (v0.71.38) Alpamys 2026-07-17 22:42:47 +0500
  • a95fedeb0e fix(cli): quote install hints so `pip install soup-cli[extra]` works on cmd.exe (v0.71.37) Alpamys 2026-07-17 20:40:54 +0500
  • 83c4c6d60b docs(readme): note the cmd.exe quoting trap in the install section Alpamys 2026-07-17 18:26:46 +0500
  • 780758a6ef test: record the MINI_MMLU coupling behind the 0.4 gate expectation (#315) Alpamys 2026-07-17 16:31:43 +0500
  • 85450309f4 docs: credit @Sanjays2402 for the benchmark gate-task fix (#315) Alpamys 2026-07-17 16:31:43 +0500
  • 3b752b7b60
         fix(eval): run built-in benchmark gate tasks (#315) Sanjay Santhanam 2026-07-17 04:30:16 -0700
  • 5271fc3b88 docs: correct the test count to the measured 16283 Alpamys 2026-07-17 12:31:37 +0500
  • 55dd6301b5 docs: v0.71.36 Data Moat II (version bump + CHANGELOG + docs) Alpamys 2026-07-17 12:03:15 +0500
  • fd51d5cfbe fix(canary): friendly "manifest not found" instead of a raw OS error Alpamys 2026-07-16 22:53:19 +0500
  • cf08cdf82f fix(gpu): measure a LOCAL checkpoint's size instead of guessing from its name Alpamys 2026-07-16 22:50:34 +0500
  • 6e718bda31 fix(canary): verdict must account for multiple comparisons (live-smoke find) Alpamys 2026-07-16 22:26:58 +0500
  • 6308038024 test(embed): de-vacuous the model_id guards — assert the SPECIFIC message Alpamys 2026-07-16 21:50:45 +0500
  • 1eaeb8e46b fix(canary): harden the secret store — 0600 + ESC-strip + entry cap Alpamys 2026-07-16 18:39:50 +0500
  • b0ac95ecdc fix(loader): replay rows bypassed vision/audio traversal protection (HIGH) Alpamys 2026-07-16 18:26:20 +0500
  • 5dd9917cee fix(v0.71.36): address ECC python-review findings Alpamys 2026-07-16 18:05:58 +0500
  • 64e80dd142 test(v0.71.36): no-top-level-torch + light-core import guards Alpamys 2026-07-16 17:43:39 +0500
  • 5ef04f16e5 feat(train): --replay / --replay-ratio passthrough (v0.71.36 Part E) Alpamys 2026-07-16 17:33:20 +0500
  • c96eee9888 feat(loader): _finalize seam + replay mix into train only (v0.71.36 Part E) Alpamys 2026-07-16 17:30:05 +0500
  • 1552e994ec feat(rehearsal): pure continual-learning replay mix (v0.71.36 Part E) Alpamys 2026-07-16 17:25:46 +0500
  • 686e885858 feat(schema): DataConfig.replay* + _validate_replay_compat (v0.71.36 Part E) Alpamys 2026-07-16 17:21:39 +0500
  • 72c1d4aa7f feat(data): soup data canary insert/check (v0.71.36 Part D) Alpamys 2026-07-16 17:17:39 +0500
  • c32fdda85a feat(canary): Secret-Sharer canaries + exposure decision rule (v0.71.36 Part D) Alpamys 2026-07-16 17:13:03 +0500
  • b25d41c95b refactor(live_eval): add compute_pair_losses; compute_eval_loss is its mean (v0.71.36 Part D) Alpamys 2026-07-16 17:09:58 +0500
  • 9f299773fb feat(data): soup data topics — BERTopic-lite coverage map (v0.71.36 Part C) Alpamys 2026-07-16 17:05:02 +0500
  • 0573009c57 feat(topics): pure k-means++ + c-TF-IDF cluster labels (v0.71.36 Part C) Alpamys 2026-07-16 16:49:54 +0500
  • 9b508d34d3 fix(cli): escape soup-cli[extra] so Rich stops eating the bracket Alpamys 2026-07-16 16:45:11 +0500
  • 6f6630997e feat(data): soup data dedup --semantic (SemDeDup) (v0.71.36 Part B) Alpamys 2026-07-16 16:28:31 +0500
  • 26f3369516 feat(semdedup): pure greedy cosine near-dup selection (v0.71.36 Part B) Alpamys 2026-07-16 16:22:40 +0500
  • e49ee0479d feat(embed): batched AutoModel encode + masked mean-pool + L2-norm (v0.71.36 Part A) Alpamys 2026-07-16 16:11:00 +0500
  • 741aba6626 test(embed): de-vacuous pooling-refusal tests + pin non-mean precedence Alpamys 2026-07-15 23:19:57 +0500
  • aed7881fc8 feat(embed): pooling gate that refuses unverified models (v0.71.36 Part A) Alpamys 2026-07-15 23:05:20 +0500
  • aa347222a6 docs: close the v0.71.35 checklist gaps (index row, toolchain, stale counts) Alpamys 2026-07-15 21:04:01 +0500
  • 7cdf32f46b test(v0.71.35): CI-proof the CLI output assertions (ANSI + wrap tolerant) Alpamys 2026-07-15 20:29:26 +0500
  • dc73e947e9 feat(compliance): init templates + soup card + soup ci init + GGUF-on-Windows (v0.71.35) Alpamys 2026-07-15 20:09:10 +0500
  • ff4c9963cb docs: v0.71.34 adapter algebra + LISA (version bump + CHANGELOG + docs) Alpamys 2026-07-15 13:39:46 +0500
  • 7e3b032ed3 test(v0.71.34): tdd-review — +36 coverage tests (factor-coeff/opt-reset/parser edges/schema matrix/CLI refusals/policy bounds), drop vacuous asserts Alpamys 2026-07-15 13:14:46 +0500
  • 1887519d8b fix(v0.71.34): security-review — H1 symlink-config reject + cwd re-validate, M1 C0/ESC strip on base_model, L1 coeff magnitude cap Alpamys 2026-07-15 13:06:19 +0500
  • 456bef9d0a fix(v0.71.34): code-review CRITICALs — LISA TrainerCallback subclass, factor-aware task-arithmetic (linear ΔW), fail-loud on no decoder layers Alpamys 2026-07-15 12:56:18 +0500
  • 57a4d329dc fix(v0.71.34): python-review — real lisa_reset_optimizer field, annotate validator, builtin generics Alpamys 2026-07-15 12:27:50 +0500
  • 2948c1e011 feat(lisa): wire LISA into SFT trainer + attach helper (v0.71.34) Alpamys 2026-07-15 12:17:28 +0500
  • 49c924df5b feat(lisa): LisaPolicy + duck-typed LisaCallback (v0.71.34) Alpamys 2026-07-15 12:15:00 +0500
  • 26d95dc233 feat(schema): LISA config fields + compat validator (v0.71.34) Alpamys 2026-07-15 12:12:04 +0500
  • e8557c260d feat(adapters): soup adapters arithmetic command (v0.71.34) Alpamys 2026-07-15 12:09:08 +0500
  • 662ecef3a7 feat(adapters): task-vector arithmetic parser + signed merge (v0.71.34) Alpamys 2026-07-15 12:04:58 +0500
  • 4e88f70691 docs: list soup draft in the serving docs-index rows Alpamys 2026-07-13 13:34:22 +0500
  • 2f79aebc5a docs(draft): v0.71.33 — soup draft, shipped as the honest measurement gate Alpamys 2026-07-13 12:46:19 +0500
  • 58ffa1a01d fix(draft): two bugs the live smoke found (v0.71.33) Alpamys 2026-07-13 12:16:51 +0500
  • 5af4e419e7 test(draft): tdd-review coverage gaps + fix a vacuous TOCTOU test (v0.71.33) Alpamys 2026-07-12 22:56:01 +0500