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): #2129 [skip ci]

Alpamysbc57893be329
diff --git a/CONTRIBUTORS.md b/CONTRIBUTORS.md
index 8214e589..686248d7 100644
--- a/CONTRIBUTORS.md
+++ b/CONTRIBUTORS.md
@@ -257,6 +257,7 @@ Listed by first contribution. PR numbers link the work.
- Added the RunPod cloud-training controller behind `soup train --cloud runpod` (#264's remaining code slice): plan-only stub, closed GPU table, token-gated submit, and a delete that only counts once a GET returns 404, run as an isolated controller (`python -I`, allow-listed environment that keeps the proxy, temporary-directory, Windows and SSH-agent settings, per-run known-hosts file, output that replaces unencodable characters). **The part worth copying is the mutation table in the description**: after the first review every unguarded check got a test that fails when it is removed, the hardening items came in the same push, and the last round pinned each forwarded variable with its own parametrized case. 73 tests in `test_runpod_hardening.py`; the 44 one-line mutations of the controller and 23 of 25 new ones fail a named test (the other two are an equivalent Windows default and an untested `hasattr` guard) ([#264](https://github.com/MakazhanAlpamys/Soup/issues/264), [#2081](https://github.com/MakazhanAlpamys/Soup/pull/2081))
- Bound single-process layer streaming to the run's recorded Hub commit (#1902, first streaming slice): tokenizer, architecture config, meta skeletons and both weight downloads share one resolved commit, and the shard and weight-copy cache slots are keyed by it as `<slug>@<commit>`. They took the review's rulings (Unsloth left unbound so its pre-quantized mapping keeps working, the wider identity work taken out into separate pull requests, the first-run re-shard stated in the changelog and guide) and closed the last gap the mutation check found: both slot tests now pin the `@` separator, so changing it to `-` or `_` fails a named test. 10 new tests plus 26 in the extended CLI test; removing any of the boundaries fails a named test ([#1902](https://github.com/MakazhanAlpamys/Soup/issues/1902), [#2094](https://github.com/MakazhanAlpamys/Soup/pull/2094))
- Closed #2119: a KTO `label` of `null`, a float, a list or an object trained as a desirable or undesirable row instead of being refused (`bool(raw_label)` made `[]` False and `["no"]` True; floats were read as `> 0`, so `0.5` was desirable and `nan` undesirable). They made the shared converter accept a bool, the six documented strings and exactly `-1`, `0` and `1`, and refuse everything else with a `ValueError` that names the label and shows the value, so `soup data validate` and `soup train --dry-run` report it before any model loads. The tests enumerate the spellings (including `1.0`, `nan`, `inf`, `-1` as a string) through the converter, the loader reason and both CLI paths, and flip the old test that pinned `None -> False`; with the file as it is on `main` 42 of 66 fail, and six source mutations (the `bool()` branch, floats as `> 0`, integral floats, any integer, `-1` inverted, a message without the value) each fail at least one named test. The explicit bool branch is redundant and survives as an equivalent mutation ([#2119](https://github.com/MakazhanAlpamys/Soup/issues/2119), [#2126](https://github.com/MakazhanAlpamys/Soup/pull/2126))
+ - Added the `qwen3.5-4b-dpo` recipe from #275: the Qwen3.5-9B DPO settings on `Qwen/Qwen3.5-4B`, text only with an explicit `modality: text`, the inherited (untuned) settings named as such in the docs and fragment. The model id is pinned on `RecipeMeta.model` and the YAML `base:` and each fails its own test; the recipe is registered in the Qwen3.5 family audit, the count moved 169 -> 170 at every site and the snapshot delta is purely additive. The anonymous `model_info` check (851bf6e8, `config.json` present, Apache-2.0) reproduces, `soup recipes verify` attaches the recipe on the real public config (adapted=8), and a tiny random-weight native Qwen3.5 text decoder is built on CPU in the tests. Ten source/test mutations each fail a named test; six default-equivalent lines (`epochs`, `batch_size`, `target_modules`, `quantization`, `dpo_beta`, `output`) survive as equivalents and are listed as such ([#275](https://github.com/MakazhanAlpamys/Soup/issues/275), [#2129](https://github.com/MakazhanAlpamys/Soup/pull/2129))
- **AmixDigital** ([@Amix29](https://github.com/Amix29))
- Added native Sesame CSM fine-tuning for transcript/audio pairs, training the backbone and depth decoder while keeping the Mimi codec frozen; documented supported inputs, refusals and the scoped Apple Silicon evidence ([#265](https://github.com/MakazhanAlpamys/Soup/issues/265), [#1769](https://github.com/MakazhanAlpamys/Soup/pull/1769)).