Skip to content

armgit / Soup

Public Imported from GitHub
0

Fine-tune LLMs from one YAML. Layer streaming trains an 8B model on a 4 GB laptop GPU.

docs(contributors): #2120 [skip ci]

Alpamyseddfe650b32f
diff --git a/CONTRIBUTORS.md b/CONTRIBUTORS.md
index 565d9ee1..83e2536d 100644
--- a/CONTRIBUTORS.md
+++ b/CONTRIBUTORS.md
@@ -668,6 +668,7 @@ Listed by first contribution. PR numbers link the work.
- A sink-based ratchet for outbound HTTP calls: the #616 guard found gates by their shape, so a differently written gate, or no gate at all, passed unchecked, and `energy.py` had its own loopback set that missed abbreviated IPv4 and `100.64.0.0/10`. They built the inverted scanner the issue asked for, and it **improved under review rather than merely passing it**: the first version saw 20 of 46 spellings of an outbound call, the two-pass one sees 45 (the missing one, an untyped client passed in as a parameter, is stated as a limit in the docstring); the table is keyed by qualified name with an exact call multiset, so a second call inside a declared function fails; and every declared gate is checked to exist and to be called (review had found four entries that named functions that do not exist). When review showed that the inert list held classes that open a connection, they removed the names rather than argue. `validate_electricity_map_endpoint` now goes through `net_guard`. 363 tests passed on the final head and 1361 on the tree merged with main, and 51 mutations were run on it, 42 caught ([#1547](https://github.com/MakazhanAlpamys/Soup/issues/1547), [#1646](https://github.com/MakazhanAlpamys/Soup/pull/1646))
- Closed #1644: the automatic evaluation that `soup train` starts called `soup eval benchmark` and `soup eval custom` with `trust_remote_code=False` whatever the run was started with, so a base that needs the opt-in trained (the user had given it) and then could not load the adapter it had just saved: `Auto-eval custom failed: ... Please pass the argument trust_remote_code=True`, exit 0, no result. The rule had been decided on the issue (the run's flag reaches the evaluation, there is no `soup.yaml` key) and they **proposed that plan on the issue and delivered exactly it**: a keyword-only parameter defaulting to `False`, passed at the one call site, reaching the two evaluation loads and nothing else, with the field refused as an unknown key at the top level and under `eval:`. **What makes it worth keeping is the test set**: a CLI test that runs the real `soup train` with only the trainer and the two evaluation commands replaced (the one that catches a call site that forgets the flag), and, after review, a test pinning the omitted-keyword default and an AST test that counts the consumers of the flag and compares the expression each receives, so a later judge or scorer load cannot inherit it by accident. Both review-round mutations that had survived (the default flipped to `True`, a third consumer) are now killed, they took the optional value check too, and they resolved the merge with #1537 and #1658 by keeping both sides and rebased the docs onto #1659's `eval` extra section. 12 new tests; 9 of 9 source mutations killed by a named test ([#1644](https://github.com/MakazhanAlpamys/Soup/issues/1644), [#1662](https://github.com/MakazhanAlpamys/Soup/pull/1662))
- Closed #1705: when remote training on a Lambda Cloud instance failed, the controller reported only the exit code and terminated the instance, so the error log was gone with it. They tee'd the remote run's output to a log under `set -euo pipefail` (so the exit code survives the pipe), and on a non-zero exit the controller fetches the tail over SSH (50 lines, 50 KB, 15 s) and prints it before terminating, with ESC, BEL, DEL and C1 bytes stripped and a fetch that can only print one warning line. **The part worth copying is how they closed the review findings**: the first request was a weakened assertion on the tail command, which they replaced with an exact one whose path is read out of the user data; the second was a survivor in which removing `capture_output` from the fetch left every test green while raw remote escape bytes went straight to the terminal, and they pinned that argument, `check=False`, the whole ssh argument list and a multibyte character split by the byte cap. The PR says plainly that no real Lambda instance was involved ([#1705](https://github.com/MakazhanAlpamys/Soup/issues/1705), [#1774](https://github.com/MakazhanAlpamys/Soup/pull/1774))
+ - Documented the accepted and refused preference row layouts for DPO and KTO (`docs/data.md`) as tables the tests read back: every documented row runs through the converter, so a refusal message that changes or a row that stops converting fails a test instead of leaving the page wrong; after review the Detection prose was brought in line with #2093 and the deletion check was rewritten to run the real count code ([#2086](https://github.com/MakazhanAlpamys/Soup/issues/2086), [#2120](https://github.com/MakazhanAlpamys/Soup/pull/2120))
- **Harsh Raj Singhania** ([@HarshRajSinghania](https://github.com/HarshRajSinghania))
- First contribution. With six of seven `[train]` packages installed, `soup doctor` announced `Training stack not installed` and recommended `pip install "soup-cli[train]"` — a command that is a no-op for the six already satisfied. `group_missing` flipped on the first absent member, so **the usual broken install, one missing wheel, was reported as no install at all**. A partial stack now names what is missing. The same patch gates the NVIDIA wheel-tag hint on torch actually being absent: with torch present on a CUDA box that hint explains how to obtain a CUDA build of a package you already have, and it was printing anyway. **What decided this over a competing patch was the test.** The wheel-tag gate had a mutation that survived on the earlier head — dropping `and torch_missing` changed nothing any test could see — and they added `test_doctor_nvidia_partial_stack_with_torch_present`, after which that mutation fails exactly one test and no others. The assertion is `"index-url" not in out` rather than a specific sentence, which catches the hint appearing in either `issues` or `fix_pre`; a sentence match would have missed the duplicate line. They also took a consolidation request seriously when arguing that the original patch already passed its own tests would have been easier ([#875](https://github.com/MakazhanAlpamys/Soup/issues/875), [#884](https://github.com/MakazhanAlpamys/Soup/pull/884))