library_name: transformers
datasets:
- kp7742/YALM-pretrain6-62M
language:
- en
- hi
pipeline_tag: text-generation
tags:
- pt
- yalm
YALM-130M
YALM (Yet Another Language Model) is a family of an experimental small language models developed through my ongoing exploration of language modeling and LLM architectures.
YALM-130M is the second model in this series. This model is trained on a diverse corpus of English, Hindi, Math, and Python Code to test its capacity for multi-lingual and technical reasoning.
Model Overview:
- Architecture: Llama
- Pretraining steps: 40k
- Pretraining tokens: 42B
- Precision: bfloat16
- Number of Parameters: 130M
- Number of Paramaters (Non-Embedding): 113M
- Number of Layers: 16
- Number of Attention Heads (GQA): 16 for Q and 2 for KV
- Context Length: 2048
Usage
>>> from transformers import AutoTokenizer, AutoModelForCausalLM
>>> tokenizer = AutoTokenizer.from_pretrained("kp7742/YALM-130M")
>>> model = AutoModelForCausalLM.from_pretrained("kp7742/YALM-130M")
>>> inputs = tokenizer("Hey how are you doing?", return_tensors="pt")
>>> out = model.generate(**inputs, max_new_tokens=100)
>>> print(tokenizer.batch_decode(out))
Training
Data
This model is pre-trained on YALM-pretrain6-62M
Hyperparameters
- learning_rate: 6e-3
- train_batch_size: 16
- eval_batch_size: 16
- distributed_type: multi-GPU DDP
- num_devices: 4
- gradient_accumulation_steps: 8
- total_train_batch_size: 512
- total_eval_batch_size: 64
- optimizer: AdamW with betas=(0.9, 0.95) and epsilon=1e-08
- lr_scheduler_type: warmup_stable_decay
- lr_scheduler_warmup_steps: 4000
- training_steps: 40000
Hardware
- GPUs: 4 x RTX 5090
Framework versions
- Transformers 4.56.2
- Pytorch 2.8.0+cu128
- Datasets 4.1.1
- Tokenizers 0.22.1
Evaluation
All evaluations are zero-shot unless stated otherwise, and I used lighteval to run them.
It achieves the following results on the test set:
- Loss: 2.46
- Perplexity: 11.66
Base pre-trained model
| Metrics | YALM-130M | YALM-80M |
|---|---|---|
| MMLU (cloze) | 27.98 | 27.33 |
| MMLU Pro | 11.38 | 8.72 |
| BBH (5-shot) | 11.59 | 12.61 |
| ARC (Average) | 33.50 | 29.87 |
| HellaSwag | 34.08 | 32.16 |
| PIQA | 62.40 | 62.89 |
| SCIQ | 70.00 | 69.50 |
| CommonsenseQA | 28.75 | 28.75 |
| Winogrande | 50.28 | 50.59 |
| OpenBookQA | 31.00 | 29.60 |
| TruthfulQA | 21.71 | 22.78 |
| TriviaQA | 0.18 | 0.17 |
| GSM8K (5-shot) | 1.06 | 0.83 |
Limitations
YALM models primarily understand and generate content in English and Hindi. They can produce text on a variety of topics but as world knowledge is limited, the generated content may not always be factually accurate, logically consistent, or free from biases present in the training data.