GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Ekansh16
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Ekansh16
Model Tree
mistralai/Mistral-7B-Instruct-v0.2
Input Modalities
Output Modalities
Supported Functionality
Try it live: Hinglish Song Generator Space
Sample output for prompt: "A clothing brand called DesiThreads launching their festive Diwali collection on October 20th for young Indians aged 18-35."
text
Verse 1:Lights glow kar rahi hain har gali aur bazaar,DesiThreads leke aaya naya fashion ka tyohaar.October 20th ko khulega humne apna darwaaza,Pura India karenge hum style ke hi baahaar. Chorus:Yeh desi vibes, yeh modern swag apna,Festive season mein pura magic dikha apna.Dil ki hai yaari aur sapnon ki udan,Let's celebrate Diwali with DesiThreads ki shaan. Verse 2:Traditional designs ka naya twist laaye hain yaar,Cuts hain trendy aur fit hai ekdum tight,Zindagi ke safar mein ab look tera bright. Outro:Hausla buland hai aur dil mein hai josh,DesiThreads ke sang aaj manao ye josh.python
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfigimport torch model_id = "Ekansh16/hinglish-song-generator-merged" # Load with 4-bit quantization (recommended — saves memory)bnb_config = BitsAndBytesConfig( load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_compute_dtype=torch.float16, bnb_4bit_use_double_quant=True,) tokenizer = AutoTokenizer.from_pretrained(model_id)tokenizer.pad_token = tokenizer.eos_token model = AutoModelForCausalLM.from_pretrained( model_id, quantization_config=bnb_config, device_map="auto",)model.eval()python
def generate_song(context: str) -> str: # Use exact Mistral instruction format prompt = f"<s>[INST] {context} [/INST]" inputs = tokenizer(prompt, return_tensors="pt").to(model.device) with torch.no_grad(): outputs = model.generate( **inputs, max_new_tokens=400, temperature=0.8, top_p=0.92, top_k=50, do_sample=True, repetition_penalty=1.2, pad_token_id=tokenizer.eos_token_id, ) input_len = inputs["input_ids"].shape[1] return tokenizer.decode(outputs[0][input_len:], skip_special_tokens=True) # Examplecontext = """A startup called ZapCart is launching their electric scooter on August 1st.Target audience is young Indians aged 18-30.Vibe should be energetic and aspirational."""print(generate_song(context))Write rich, descriptive prompts for best results. Include:
Good prompt example:
text
A wedding sangeet ceremony. Groom from Delhi, bride from Jaipur.Fun, filmy, and emotional Bollywood style.Poor prompt example:
text
Make a song5,952 training examples combining two sources:
| Source | Examples | Purpose |
|---|---|---|
| Retro Bollywood lyrics dataset (cleaned) | 3,613 | Teaches Hinglish vocabulary, rhyme patterns, and song structure |
| Synthetically generated promotional examples (Gemini API) | 1,000 × 3 oversampled | Teaches context-aware generation for brands, events, and promotions |
The synthetic dataset was generated using Gemini API with 29 varied context templates covering brand launches, weddings, festivals, college events, fitness, motivational themes, and travel.
Hardware: Kaggle T4 GPU (free tier)
Fine-tuning configuration:
python
# QLoRA settingsload_in_4bit = Truebnb_4bit_quant_type = "nf4"bnb_4bit_compute_dtype = torch.float16bnb_4bit_use_double_quant = True # LoRA settingslora_r = 16lora_alpha = 32lora_dropout = 0.05target_modules = ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"] # Training settingsnum_epochs = 3batch_size = 2gradient_accumulation_steps = 8 # effective batch = 16learning_rate = 2e-4optimizer = "paged_adamw_8bit"lr_scheduler = "cosine"max_seq_length = 512Training results:
| Step | Train Loss | Val Loss |
|---|---|---|
| 100 | 1.4139 | 1.4801 |
| 400 | 1.0916 | 1.1241 |
| 700 | 0.8274 | 0.7981 |
| 1000 | 0.2149 | 0.6154 |
| 1100 | 0.1879 | 0.6069 |
The model always generates in this structure:
text
Verse 1:[4 lines, rhyming] Chorus:[4 catchy, repeatable lines] Verse 2:[4 lines, rhyming] Outro:[2 closing lines]Hinglish style characteristics:
✅ Brand promotional songs and jingles
✅ Wedding sangeet and event songs
✅ Festival and Diwali/Holi themed songs
✅ Motivational and sports anthems
✅ Social media reel background lyrics
✅ Birthday and farewell songs
❌ Pure Hindi poetry or Devanagari output
❌ Classical music or formal compositions
❌ Languages other than Hinglish/English
If you use this model in your work, please cite:
text
@misc{hinglish-song-generator-2025, author = {Ekansh Patidar}, title = {Hinglish Song Generator: Fine-tuned Mistral 7B for Hinglish Lyric Generation}, year = {2026}, publisher = {HuggingFace}, url = {https://huggingface.co/Ekansh16/hinglish-song-generator-merged}}Verse 1:Lights glow kar rahi hain har gali aur bazaar,DesiThreads leke aaya naya fashion ka tyohaar.October 20th ko khulega humne apna darwaaza,Pura India karenge hum style ke hi baahaar.Chorus:Yeh desi vibes, yeh modern swag apna,Festive season mein pura magic dikha apna.Dil ki hai yaari aur sapnon ki udan,Let's celebrate Diwali with DesiThreads ki shaan.Verse 2:Traditional designs ka naya twist laaye hain yaar,Cuts hain trendy aur fit hai ekdum tight,Zindagi ke safar mein ab look tera bright.Outro:Hausla buland hai aur dil mein hai josh,DesiThreads ke sang aaj manao ye josh.from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfigimport torch model_id = "Ekansh16/hinglish-song-generator-merged" # Load with 4-bit quantization (recommended — saves memory)bnb_config = BitsAndBytesConfig( load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_compute_dtype=torch.float16, bnb_4bit_use_double_quant=True,) tokenizer = AutoTokenizer.from_pretrained(model_id)tokenizer.pad_token = tokenizer.eos_token model = AutoModelForCausalLM.from_pretrained( model_id, quantization_config=bnb_config, device_map="auto",)model.eval()def generate_song(context: str) -> str: # Use exact Mistral instruction format prompt = f"<s>[INST] {context} [/INST]" inputs = tokenizer(prompt, return_tensors="pt").to(model.device) with torch.no_grad(): outputs = model.generate( **inputs, max_new_tokens=400, temperature=0.8, top_p=0.92, top_k=50, do_sample=True, repetition_penalty=1.2, pad_token_id=tokenizer.eos_token_id, ) input_len = inputs["input_ids"].shape[1] return tokenizer.decode(outputs[0][input_len:], skip_special_tokens=True) # Examplecontext = """A startup called ZapCart is launching their electric scooter on August 1st.Target audience is young Indians aged 18-30.Vibe should be energetic and aspirational."""print(generate_song(context))# QLoRA settingsload_in_4bit = Truebnb_4bit_quant_type = "nf4"bnb_4bit_compute_dtype = torch.float16bnb_4bit_use_double_quant = True # LoRA settingslora_r = 16lora_alpha = 32lora_dropout = 0.05target_modules = ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"] # Training settingsnum_epochs = 3batch_size = 2gradient_accumulation_steps = 8 # effective batch = 16learning_rate = 2e-4optimizer = "paged_adamw_8bit"lr_scheduler = "cosine"max_seq_length = 512Verse 1:[4 lines, rhyming]Chorus:[4 catchy, repeatable lines]Verse 2:[4 lines, rhyming]Outro:[2 closing lines]@misc{hinglish-song-generator-2025, author = {Ekansh Patidar}, title = {Hinglish Song Generator: Fine-tuned Mistral 7B for Hinglish Lyric Generation}, year = {2026}, publisher = {HuggingFace}, url = {https://huggingface.co/Ekansh16/hinglish-song-generator-merged}}