Fine-tune LLMs from one YAML. Layer streaming trains an 8B model on a 4 GB laptop GPU.
Python cliconsumer-gpudpofine-tuningggufhuggingfacellmllmopslocal-ailocal-llmloralow-vrammachine-learningollamapeftpythonpytorchqlorasfttransformers
docs(contributors): #2113 [skip ci]
Alpamysb93dd7e9238f
diff --git a/CONTRIBUTORS.md b/CONTRIBUTORS.mdindex 590651cc..d873f89b 100644--- a/CONTRIBUTORS.md+++ b/CONTRIBUTORS.md@@ -630,6 +630,7 @@ Listed by first contribution. PR numbers link the work.- Took the first of the six saturated bundled suites in the #1192 series off the 1.000 rail: `mini_arithmetic` replaced 12 single-step rows with 16 multi-step or large-operand ones, bumped `BUNDLED_SCORER_REVISION` to 4 and explained why the fingerprint correctly did not move, and published a record that says it is not the 7B reference and that half the strong model's misses are truncation (Qwen2.5-1.5B 0.800, SmolLM2-135M 0.375). After review the test now recomputes every hard key from its own question, which closed two surviving mutations (a wrong question with its key kept); twelve mutations killed, new file fails 5 of 7 on clean main ([#1192](https://github.com/MakazhanAlpamys/Soup/issues/1192), [#1462](https://github.com/MakazhanAlpamys/Soup/pull/1462))- `soup data augment` wrote rows with an empty rewritten field and `soup distill-prompt` dropped every failed row and wrote an empty file, both exiting 0 while every provider call failed, because neither built its provider with `raise_on_error`; they fixed the cause in both builders, counted every call, treated an empty reply as a failure, exited 1 without a file when nothing was usable and reported the failure count on a partial run, and replaced the five tests that had encoded the defect with ones asserting the raised error and the exact count. After review they escaped provider error text before Rich prints it (an error containing `[/x]` crashed the command, after the file was written on the partial path), tightened five `!= 0` assertions to `== 1`, and covered translate, style and all three escape sites. 16 of 18 new tests fail on clean main, fifteen mutations killed ([#1274](https://github.com/MakazhanAlpamys/Soup/issues/1274), [#1463](https://github.com/MakazhanAlpamys/Soup/pull/1463))- `mini_common_sense` scored 1.000 on the strong reference, so the forgetting gate could not see a regression on it, and 14 of its 24 keys were B. They grew the fixture to 40 rows from a measured candidate pool, rebalanced the key so no letter exceeds 30%, bumped the scorer revision with the fingerprint reasoning written next to it, and recorded the headroom and the chance level in a benchmark record. **When review showed that five key and wording mutations left every test green, they added the two tests that pin each key to its option text and each hard stem to its question, verbatim and in one push**; all five mutations fail now, reverting the fixture fails 8 tests, and 861 pass around the gate ([#1192](https://github.com/MakazhanAlpamys/Soup/issues/1192), [#1673](https://github.com/MakazhanAlpamys/Soup/pull/1673))+ - `mini_mmlu` scored 1.000 on the strong reference and 12 of its 26 keys were B, so the gate could not see a regression on it. They grew the fixture to 40 rows from two measured candidate pools, rebalanced the key so no letter exceeds 30%, bumped the scorer revision with the fingerprint reasoning next to #1958's note, and moved two noise-floor tests off the lopsided key without weakening an assertion; the record states that no 7B model was measured ([#1192](https://github.com/MakazhanAlpamys/Soup/issues/1192), [#2113](https://github.com/MakazhanAlpamys/Soup/pull/2113))- **Aditya Pratap Singh** ([@akkupratap323](https://github.com/akkupratap323))- Closed a silent pass in the PR gate `soup ci init` writes: `soup data validate` exited 0 at "0/5 rows valid", so a dataset the loader would drop entirely sailed through CI. Zero usable rows now fails with exit 2 and a reason, while a partial drop still passes as it did — **the change is exactly as narrow as the issue asked**, and it reuses the pass-0 / failed-gate-2 / input-error-1 shape `soup ship` and `soup data lint` already use rather than adding a sixth convention. An opt-in `--min-valid-fraction` adds a stricter gate, with the equal-to-threshold case pinned by a test. Every test kills its mutation, including over-broadening to "any dropped row", which fails four tests. A first contribution ([#811](https://github.com/MakazhanAlpamys/Soup/issues/811), [#858](https://github.com/MakazhanAlpamys/Soup/pull/858))