159 lines
5.5 KiB
Markdown
159 lines
5.5 KiB
Markdown
# 🎙️ Alex-Tech-Recruiter-v1 (Llama 3.1 8B)
|
|
|
|
<p align="center">
|
|
<strong>A fine-tuned Llama 3.1 8B Instruct model that simulates a professional technical recruiter named "Alex" for realistic multi-turn interview roleplay.</strong>
|
|
</p>
|
|
|
|
---
|
|
|
|
## 📋 Model Overview
|
|
|
|
| Key | Value |
|
|
|---|---|
|
|
| **License** | Llama 3.1 Community License |
|
|
| **Base Model** | `unsloth/llama-3-8b-Instruct-bnb-4bit` |
|
|
| **Finetuning Method** | QLoRA (Quantized Low-Rank Adaptation) via Unsloth |
|
|
| **Language** | English |
|
|
| **Pipeline Tag** | `text-generation` |
|
|
|
|
### Tags
|
|
`llama-3.1` · `recruiter` · `hr-tech` · `interview-simulation` · `roleplay` · `qlora` · `finetuned`
|
|
|
|
---
|
|
|
|
## 📖 Model Description
|
|
|
|
**Alex-Tech-Recruiter-v3** is a fine-tuned version of Llama 3.1 8B Instruct, designed to simulate a highly realistic, professional technical recruiter named **"Alex."** It specializes in conducting multi-turn voice-style interviews grounded in specific candidate CVs and Job Descriptions (JDs).
|
|
|
|
### Model Sources
|
|
|
|
| Resource | Link |
|
|
|---|---|
|
|
| **Repository** | [https://huggingface.co/oussema2021/fintuned_v3_AiRecruter] |
|
|
| **Dataset** | [pending owner consent... ] |
|
|
| **Base Model** | [unsloth/llama-3-8b-Instruct-bnb-4bit](https://huggingface.co/unsloth/llama-3-8b-Instruct-bnb-4bit) |
|
|
|
|
---
|
|
|
|
## 🚀 Uses
|
|
|
|
### Direct Use
|
|
|
|
This model is intended for:
|
|
|
|
- **Interview Simulation** — Acting as a practice partner for candidates preparing for technical interviews. Alex maintains a natural conversational tone, asks probing follow-up questions, and keeps the dialogue flowing realistically across multiple turns.
|
|
|
|
- **Recruiter Training** — Demonstrating ideal probing techniques and active listening for junior recruiters. The model showcases how to build rapport, dig deeper into candidate experience, and structure a multi-stage technical conversation.
|
|
|
|
- **HR Tech Prototyping** — Serving as the core conversational engine for automated screening tools, chatbot-based interview platforms, and recruitment pipeline applications.
|
|
|
|
### Out-of-Scope Use
|
|
|
|
> ⚠️ This model should **not** be used for making hiring decisions autonomously. It is a simulation tool and may hallucinate details not present in the provided context. Always pair it with human oversight in any real-world hiring workflow.
|
|
|
|
---
|
|
|
|
## ⚠️ Bias, Risks, and Limitations
|
|
|
|
| Limitation | Description |
|
|
|---|---|
|
|
| **Hallucination** | The model may occasionally hallucinate details if the CV/JD context is missing or ambiguous. Always provide structured, up-to-date context for best results. |
|
|
| **Length Bias** | While trained on full transcripts, extremely long conversations (>20 turns) may see a degradation in context retention. Consider summarizing earlier turns for extended sessions. |
|
|
| **Persona Rigidity** | The model is heavily fine-tuned on the "Alex" persona. Switching to a different persona via system prompt may require additional few-shot examples to override the training. |
|
|
|
|
---
|
|
|
|
## 🛠️ How to Get Started
|
|
|
|
### Installation
|
|
|
|
```bash
|
|
pip install unsloth transformers accelerate bitsandbytes
|
|
```
|
|
|
|
### Python Code
|
|
|
|
```python
|
|
from unsloth import FastLanguageModel
|
|
import torch
|
|
|
|
# Load the model and tokenizer
|
|
model, tokenizer = FastLanguageModel.from_pretrained(
|
|
model_name = "your-username/alex-tech-recruiter-v1", # Replace with your HF ID
|
|
max_seq_length = 4096,
|
|
dtype = None,
|
|
load_in_4bit = True,
|
|
)
|
|
|
|
# Enable faster inference
|
|
FastLanguageModel.for_inference(model)
|
|
|
|
# Define the input data
|
|
candidate_name = "John Doe"
|
|
job_role = "Senior Data Engineer"
|
|
job_description = """
|
|
We are looking for a Senior Data Engineer to build scalable pipelines...
|
|
"""
|
|
candidate_cv = """
|
|
John Doe
|
|
Senior Data Engineer at Google (2020-Present)
|
|
Skills: Python, Spark, Kubernetes...
|
|
Experience: Led migration of legacy data warehouse to BigQuery...
|
|
"""
|
|
|
|
# Format the System Prompt
|
|
system_prompt = f"""You are Alex, a warm and professional senior technical recruiter \
|
|
conducting a live voice interview with {candidate_name} for the role of {job_role}.
|
|
|
|
This is a natural spoken conversation, not a written exchange. Speak as you would in a real interview.
|
|
|
|
## Conversation Flow
|
|
**Opening (first turn only):**
|
|
- Greet the candidate warmly.
|
|
- Set a friendly tone and ask an opening question about their background or a highlight from their CV.
|
|
|
|
**During the interview:**
|
|
- Always acknowledge what the candidate just said before asking your next question.
|
|
- Ask ONE focused follow-up question that digs deeper into their experience.
|
|
|
|
**Closing:**
|
|
- Wrap up the interview naturally, thank the candidate, and outline next steps.
|
|
"""
|
|
|
|
# Build the conversation messages
|
|
messages = [
|
|
{"role": "system", "content": system_prompt},
|
|
{"role": "user", "content": candidate_cv},
|
|
]
|
|
|
|
# Generate the first recruiter response
|
|
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")
|
|
outputs = model.generate(
|
|
input_ids=inputs,
|
|
max_new_tokens=512,
|
|
temperature=0.7,
|
|
top_p=0.9,
|
|
do_sample=True,
|
|
)
|
|
response = tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True)
|
|
print(response)
|
|
```
|
|
|
|
---
|
|
|
|
## 🏗️ Training Details
|
|
|
|
- **Finetuning Method**: QLoRA (Quantized Low-Rank Adaptation)
|
|
- **Framework**: [Unsloth](https://github.com/unslothai/unsloth)
|
|
- **Precision**: 4-bit quantization (BitsAndBytes)
|
|
- **Max Sequence Length**: 4096 tokens
|
|
|
|
---
|
|
|
|
## 📄 License
|
|
|
|
This model is released under the **Llama 3.1 Community License**. Please review the full license terms before using this model in production or commercial applications.
|
|
|
|
---
|
|
|
|
*Developed with ❤️ using [] and Hugging Face Transformers.* |