Serenade — SmolLM2 addition pilot
I fine-tuned SmolLM2-135M-Instruct locally with LoRA on a GTX 1650 Max-Q with 4 GB VRAM. I developed the scripts and experiment.
This repository contains the merged model. It is a fine-tune of Hugging Face's pretrained model, separate from my native Serenade model trained from random weights.
Training
- 1,000 generated addition examples.
- Operands from 0 to 49, presented in canonical order.
- 100 validation and 100 held-out test examples.
- Unordered addition facts are disjoint across these splits.
- 100 optimizer steps; batch 1; gradient accumulation 8.
- AdamW; learning rate 0.0002; float32.
- LoRA rank 8, alpha 16, dropout 0.05.
- Targets: q_proj and v_proj.
- Trainable parameters: 460,800.
- Validation loss: 2.2458 to 0.2851.
Local evaluation
Both models used greedy generation, a 12-token output limit, and the same 100 held-out questions.
| Scoring | Original model | Fine-tuned model |
|---|---|---|
| Complete integer-only answer | 0/100 | 51/100 |
| Complete number or matching equation | 46/100 | 51/100 |
The second scoring rule was added after inspecting outputs. It is a diagnostic analysis, not a preregistered benchmark. The original strict score penalizes correct equation-form answers.
Incomplete responses: original model 7/100; fine-tuned model 0/100.
Raw outputs and training artifacts are in experiment/.
Limitations
This small, single-seed experiment mainly demonstrates improved answer formatting. The observed arithmetic gain is limited. It does not establish general reasoning ability, reliable mathematics, or performance outside the tested number range.
GSM8K evaluation has not been completed. These results are not GSM8K scores. Course acceptance and certification are pending.
Use the model's existing chat template. This is an experimental model and should not be relied upon for accurate calculations.