⚠️ IMPORTANT: This repository contains ONLY the LoRA adapter weights.
You must load the base model Qwen/Qwen3-0.6B separately.
See the Usage section below for complete loading instructions.

Model Details

Base Model: Qwen/Qwen3-0.6B
Fine-tuning Method: LoRA (Low-Rank Adaptation)
Training Dataset: sandeeppanem/resume-json-extraction-5k
Training Examples: 4,879 resumes with structured JSON outputs
Training Framework: TRL (SFTTrainer) with PEFT

Training Configuration

LoRA Parameters

LoRA Rank (r): 8
LoRA Alpha: 16
LoRA Dropout: 0.05
Target Modules: q_proj, v_proj, k_proj, o_proj
Bias: none (not trained)
Trainable Parameters: ~2.3M (0.38% of total 598M parameters)

Training Hyperparameters

Learning Rate: 1e-4
Batch Size: 2 per device
Gradient Accumulation Steps: 8
Effective Batch Size: 16 (2 × 8)
Epochs: 3
Max Sequence Length: 1536
Optimizer: adamw_torch
Weight Decay: 0.01
Learning Rate Scheduler: cosine
Warmup Steps: 100
Gradient Checkpointing: Enabled (for memory efficiency)
Mixed Precision: fp16 (float16)

Usage

Load Model

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3-0.6B",
    torch_dtype=torch.float16,
    device_map="auto",
    trust_remote_code=True
)

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "sandeeppanem/qwen3-0.6b-resume-json")
model.eval()  # Set to evaluation mode

tokenizer = AutoTokenizer.from_pretrained("sandeeppanem/qwen3-0.6b-resume-json", trust_remote_code=True)

# Set padding token if not present
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token
    tokenizer.pad_token_id = tokenizer.eos_token_id

Inference

python

import torch
import json

# Prepare input
resume_text = """Your resume text here..."""

messages = [
    {
        "role": "system",
        "content": "You are an expert resume parser. Extract structured information from resumes and return ONLY valid JSON. Do not include explanations or extra text."
    },
    {
        "role": "user",
        "content": f"Resume:\n{resume_text}"
    }
]

# Apply chat template
prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False
)

# Tokenize and generate
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=512,
        do_sample=False,  # IMPORTANT: Use deterministic generation for JSON
        pad_token_id=tokenizer.eos_token_id
    )

# Extract only the assistant's response (after the input prompt)
assistant_response = tokenizer.decode(
    outputs[0][inputs["input_ids"].shape[-1]:],
    skip_special_tokens=True
).strip()

# Parse JSON
try:
    parsed_json = json.loads(assistant_response)
    print("✓ Valid JSON")
    print(json.dumps(parsed_json, indent=2))
except json.JSONDecodeError as e:
    print(f"⚠️ Invalid JSON: {e}")
    print("Raw response:", assistant_response)

Model Performance

This model was fine-tuned on 4,879 resume examples and is optimized for:

Extracting structured information from unstructured resume text
Generating valid JSON output without explanations
Handling diverse resume formats and job titles

Limitations

The model is fine-tuned specifically for resume parsing tasks
Performance may vary with resumes in languages other than English
Very long resumes (>1536 tokens) may be truncated
The model requires the base Qwen3-0.6B model to function

Citation

If you use this model, please cite:

bibtex

@misc{qwen3-resume-json,
  title={Qwen3-0.6B Resume JSON Extraction Model},
  author={Sandeep Panem},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/sandeeppanem/qwen3-0.6b-resume-json}}
}

License

This model is licensed under Apache 2.0, same as the base Qwen3-0.6B model.