| # Smol-d1 spot-check | |
| A bounded reproducibility probe for [FlameF0X/smol-d1](https://huggingface.co/FlameF0X/smol-d1) and the stated [SmolLM2-135M](https://huggingface.co/HuggingFaceTB/SmolLM2-135M) base. It is not a benchmark or quality claim. | |
| ## Finding | |
| The base checkpoint loaded as `LlamaForCausalLM` with 134,515,008 measured parameters. A fixed-seed (21), three-example greedy generation control produced short continuations. The adapter's structured decision API was not executed because Unsloth is not installed in the test environment; accordingly, no adapter accuracy or comparison is claimed. See `results.json` for outputs and limitations. | |
| ## Reproduce | |
| Download both repositories locally and adjust the paths at the top of `smol_d1_comparison.py` if necessary. Install the dependencies supported by Smol-d1's card, including Unsloth, to run the adapter branch; absent it, the script records that limitation and runs only the base continuation control. Run: | |
| ```bash | |
| python smol_d1_comparison.py | |
| ``` | |