GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
M-amir1998
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
M-amir1998
Model Tree
HuggingFaceTB/SmolLM2-135M-Instruct
Input Modalities
Output Modalities
Supported Functionality
base -> this adapter. Adapters listed before this one must be merged into the
base model, in order, before applying this adapter.
Dataset data/arith-v1, split sft_train_3k
(fraction 1.0): synthetic arithmetic problems with exact
integer answers. See the dataset card in the repository.
yaml
epochs: 32.0max_steps: nulllearning_rate: 0.0002batch_size: 16gradient_accumulation_steps: 1warmup_ratio: 0.05weight_decay: 0.0lr_scheduler: linearmax_grad_norm: 1.0max_length: 256logging_steps: 5save_steps: nullsave_total_limit: 1gradient_checkpointing: falsefull_finetune: falselora: r: 16 alpha: 32 dropout: 0.05 target_modules: all-linearSeed 42; trained on cuda in
3542.8 s; 4884480 trainable
parameters (3.50%).
Greedy decoding on split test
(500 problems), exact-match on the final answer line.
| Metric | Value |
|---|---|
| Accuracy | 0.686 [0.644, 0.725] |
| Format compliance | 1.000 |
| Preference accuracy | 0.992 |
Research artifact for studying post-training methods on a controlled task. It is specialised to short integer arithmetic in one prompt format, was trained on a small model, and should not be used for anything requiring reliable arithmetic.