GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
shubham4saurav
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
shubham4saurav
Model Tree
Qwen/Qwen3-4B
Input Modalities
Output Modalities
Supported Functionality
python
from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizerbase = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-4B", torch_dtype="bfloat16")model = PeftModel.from_pretrained(base, "shubham4saurav/qwen3-4b-gita-lora")tok = AutoTokenizer.from_pretrained("shubham4saurav/qwen3-4b-gita-lora")msgs = [{"role": "user", "content": "Recite Bhagavad Gita 2.47."}]prompt = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True,enable_thinking=False)out = model.generate(**tok(prompt, return_tensors="pt"), max_new_tokens=256)print(tok.decode(out[0], skip_special_tokens=True))
| Base | unsloth/Qwen3-4B-unsloth-bnb-4bit |
| Method | LoRA (Unsloth + TRL SFTTrainer) |
| Rank / alpha / dropout | 16 / 16 / 0 |
| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
| Max sequence length | 2048 |
| Epochs | 3 (945 steps) |
| LR / batch / grad-accum | 2e-4 / 2 / 4 |
| Final train loss | 1.10 |
| Data | 2,520 instruction examples from 630 Gita verses |
Data source: JDhruv14/Bhagavad-Gita_Dataset
(701 verses). Each verse expands into 4 tasks — explanation, translation, quote,
transliteration — with randomised prompt templates.
Every training target is short and templated (verse + a one-line translation), even for "explain" prompts, so the adapter makes the model answer tersely and stop early. On a 20-prompt held-out eval:
| avg words / answer | chrF vs gold | |
|---|---|---|
| Base Qwen3-4B | 101 | 0.188 |
| + this adapter | 38 | 0.227 |
Higher chrF (closer to the terse gold answers) but far less elaboration. It is a learning project, not a scholarly reference — do not rely on it for authoritative interpretation of scripture.
MIT (adapter weights). The base model keeps its own license.