91 lines
2.8 KiB
Markdown
91 lines
2.8 KiB
Markdown
|
---
|
|||
|
|
license: apache-2.0
|
|||
|
|
base_model: HuggingFaceTB/SmolLM2-1.7B-Instruct
|
|||
|
|
tags:
|
|||
|
|
- text-generation-inference
|
|||
|
|
- transformers
|
|||
|
|
- lora
|
|||
|
|
- trl
|
|||
|
|
- smollm2
|
|||
|
|
- aptitude
|
|||
|
|
datasets:
|
|||
|
|
- Prathamesh25/aptitude-qa-dataset
|
|||
|
|
language:
|
|||
|
|
- en
|
|||
|
|
pipeline_tag: text-generation
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
# SmolLM2 1.7B Aptitude QA
|
|||
|
|
|
|||
|
|
This model is a fine-tuned version of `HuggingFaceTB/SmolLM2-1.7B-Instruct` optimized for technical aptitude question-answering tasks, including basic programming concepts (C, Python, loops) and fundamental AI/ML reasoning logic.
|
|||
|
|
|
|||
|
|
It was trained using LoRA parameters in pure Float16 precision on a free Google Colab T4 GPU node to handle the target dataset schema formatting seamlessly.
|
|||
|
|
|
|||
|
|
## Model Description
|
|||
|
|
|
|||
|
|
- **Developed by:** Prathamesh25
|
|||
|
|
- **Base Model:** [HuggingFaceTB/SmolLM2-1.7B-Instruct](https://huggingface.co)
|
|||
|
|
- **Dataset Used:** [Prathamesh25/aptitude-qa-dataset](https://huggingface.co)
|
|||
|
|
- **Language:** English
|
|||
|
|
- **License:** Apache 2.0
|
|||
|
|
|
|||
|
|
## Training Information & Hyperparameters
|
|||
|
|
|
|||
|
|
- **Precision:** Pure Float16 (`fp16=True`)
|
|||
|
|
- **Optimizer:** `adamw_torch`
|
|||
|
|
- **Learning Rate:** 2e-4
|
|||
|
|
- **Per Device Train Batch Size:** 4
|
|||
|
|
- **Gradient Accumulation Steps:** 4
|
|||
|
|
- **LoRA Configuration:** Rank ($r$)=16, Alpha ($lpha$)=32, Dropout=0.05
|
|||
|
|
- **Target Modules:** `q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`
|
|||
|
|
|
|||
|
|
### Loss Progression Metrics
|
|||
|
|
The training process showed highly stable optimization across epochs:
|
|||
|
|
- **Step 10:** 1.955
|
|||
|
|
- **Step 30:** 0.709
|
|||
|
|
- **Step 100:** 0.412
|
|||
|
|
- **Step 200:** 0.323
|
|||
|
|
- **Step 330:** 0.281
|
|||
|
|
|
|||
|
|
## How to Use
|
|||
|
|
|
|||
|
|
```python
|
|||
|
|
import torch
|
|||
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|||
|
|
|
|||
|
|
model_id = "Prathamesh25/smollm2-1.7b-aptitude-qa-v1"
|
|||
|
|
|
|||
|
|
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
|||
|
|
model = AutoModelForCausalLM.from_pretrained(
|
|||
|
|
model_id,
|
|||
|
|
torch_dtype=torch.float16,
|
|||
|
|
device_map="auto"
|
|||
|
|
)
|
|||
|
|
|
|||
|
|
test_prompt = [
|
|||
|
|
{
|
|||
|
|
"role": "user",
|
|||
|
|
"content": "subject: 06 technical aptitude (basic programming and aiml concepts)\nlevel: 01 basic\ncategory: loops\nquestion: what is the output of this c code: `int sum = 0; for(int i=1; i<=3; i++) { sum += i; } printf(\"%d\", sum);`?"
|
|||
|
|
}
|
|||
|
|
]
|
|||
|
|
|
|||
|
|
text_prompt = tokenizer.apply_chat_template(test_prompt, tokenize=False, add_generation_prompt=True)
|
|||
|
|
model_inputs = tokenizer([text_prompt], return_tensors="pt").to("cuda")
|
|||
|
|
input_token_len = model_inputs.input_ids.shape[1]
|
|||
|
|
|
|||
|
|
with torch.no_grad():
|
|||
|
|
generated_ids = model.generate(
|
|||
|
|
input_ids=model_inputs.input_ids,
|
|||
|
|
attention_mask=model_inputs.attention_mask,
|
|||
|
|
max_new_tokens=150,
|
|||
|
|
temperature=0.1,
|
|||
|
|
do_sample=True,
|
|||
|
|
eos_token_id=tokenizer.eos_token_id,
|
|||
|
|
pad_token_id=tokenizer.eos_token_id
|
|||
|
|
)
|
|||
|
|
|
|||
|
|
final_tokens = generated_ids[0][input_token_len:]
|
|||
|
|
print(tokenizer.decode(final_tokens, skip_special_tokens=True))
|
|||
|
|
```
|
|||
|
|
---
|