Modelpublic

shimogerald/lora_interview_coach

sourceHugging Facemitupdated 2mo agoView on Hugging Face

0likes8downloads

Model Card

Interview Coach LoRA (Qwen2.5-3B-Instruct)

LoRA adapter fine-tuned for software-engineering interview Q&A coaching.

Model Details

—Base model: unsloth/Qwen2.5-3B-Instruct
—Method: QLoRA (4-bit) + LoRA via Unsloth
—LoRA: r=16, lora_alpha=16, lora_dropout=0
—Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
—Context length: 2048
—Language: English

Training Data

Fine-tuned on `shimogerald/interview-coach-dataset` (chat messages format, ~90/10 train/val).

Intended Use

Practice / coaching-style answers to technical interview questions (APIs, systems, coding concepts, behavioral, etc.).

Limitations

—Synthetic training data may contain errors
—Not a substitute for real interview feedback
—May hallucinate technical details
—English only

How to Use

python

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="shimogerald/lora_interview_coach",
    max_seq_length=2048,
    load_in_4bit=True,
)
FastLanguageModel.for_inference(model)

messages = [{"role": "user", "content": "What is the difference between PUT and PATCH?"}]
prompt = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=256, do_sample=False)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

If loading the adapter separately fails, load the base model then attach this repo with PEFT PeftModel.from_pretrained.

Training Setup (summary)

—Optimizer: AdamW
—LR schedule: cosine with warmup
—Epochs: 3
—Framework: Unsloth + Accelerate + Transformers

This qwen2 model was trained 2x faster with Unsloth