Qwen3_4B_NCRT_Physics_12th_Finetuned API & Inference Endpoint

📌 Model Overview

Model Type: PEFT LoRA Adapter (Not Merged)
Base Model: unsloth/Qwen3-4B-Thinking-2507
Fine-Tuning Method: QLoRA (4-bit NF4 quantized base model + LoRA rank 128)
Trainable Parameters: 264,241,152 (6.16% of total 4.28B parameters)
Dataset Scope: Sample subset of Class 12 Physics topics (4,715 examples)

🎯 Intended Use & Limitations

Primary Use Cases

Assisting with high school physics conceptual questions and standard numerical problems.
Demonstrating step-by-step reasoning and physics derivations with explicit internal monologue (<think> tags).
Interactive physics tutoring on key topics represented in the sample dataset.

Limitations & Out-of-Scope

Sample Dataset, Not Comprehensive: The training data is a sample collection (4,715 examples) and does not span all chapters, subtopics, or edge cases of the complete 12th-grade NCERT curriculum.
Requires Base Model: As this repo hosts adapter weights only, it cannot be run without the base model unsloth/Qwen3-4B-Thinking-2507.
Numerical Verification: Complex multi-step calculations should always be cross-checked against standard reference textbooks.

📊 Dataset Details

Source: [KadamParth/NCERT_Physics_12th]
Nature of Data: Sample dataset containing 4,715 Physics question-answer pairs with step-by-step explanations.

Conversation Schema

text
<|im_start|>system
Follow these steps to answer the user's question:
1. First, reason step-by-step inside <think> tags. This is your internal monologue.
2. After the </think> tag, provide a complete, helpful, and polished answer for the user that includes both a direct answer and an explanation.<|im_end|>
<|im_start|>user
{Question}<|im_end|>
<|im_start|>assistant
<think>
{Step-by-step reasoning and physics derivations}
</think>

Answer: {Direct Answer}
Explanation: {Detailed Explanation}<|im_end|>

📈 Training & Validation Loss

Table with columns: Step, Training Loss, Validation Loss
Step	Training Loss	Validation Loss
10	0.8160	0.5512
20	0.5019	0.4574
30	0.4581	0.4241
40	0.4187	0.4070
50	0.4173	0.3952
60	0.4043

Final Training Loss: 0.4640
Final Validation Loss: 0.3805

🚀 How to Run Inference

1. Using Unsloth (Loading Base Model + Adapter)

python
from unsloth import FastLanguageModel
from transformers import TextStreamer

# Load the base model with adapter weights
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "Rakesh7n/Qwen3_4B_NCRT_Physics_12th_Finetuned", # points to this LoRA adapter repo
    max_seq_length = 2048,
    load_in_4bit = True,
)
FastLanguageModel.for_inference(model)

# Prepare prompt
messages = [
    {
        "role": "user",
        "content": "If 10^10 electrons move from one body to another every second, how long will it take to transfer 1 C of charge?"
    }
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize = False,
    add_generation_prompt = True,
)

inputs = tokenizer(text, return_tensors = "pt").to("cuda")
streamer = TextStreamer(tokenizer, skip_prompt = False)

# Generate with reasoning
_ = model.generate(
    **inputs,
    max_new_tokens = 2048,
    temperature = 0.7,
    top_p = 0.8,
    top_k = 20,
    streamer = streamer,
)

2. Using Standard Hugging Face `transformers` & `peft`

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

base_model_id = "unsloth/Qwen3-4B-Thinking-2507"
adapter_model_id = "Rakesh7n/Qwen3_4B_NCRT_Physics_12th_Finetuned"

tokenizer = AutoTokenizer.from_pretrained(base_model_id)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.float16,
    device_map="auto"
)

# Load PEFT LoRA adapter on top of base model
model = PeftModel.from_pretrained(base_model, adapter_model_id)

messages = [
    {
        "role": "system",
        "content": "Follow these steps to answer the user's question:\n1. First, reason step-by-step inside <think> tags. This is your internal monologue.\n2. After the </think> tag, provide a complete, helpful, and polished answer for the user that includes both a direct answer and an explanation."
    },
    {
        "role": "user",
        "content": "What is the formula for electric flux according to Gauss's Law?"
    }
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

outputs = model.generate(
    **inputs,
    max_new_tokens=1024,
    temperature=0.7,
    top_p=0.8,
    do_sample=True
)

print(tokenizer.decode(outputs[0], skip_special_tokens=False))

📝 Example Output

User Prompt:

If $1010$ electrons move from one body to another every second, how long will it take to transfer $1 C$ of charge?

Model Output:

text
<think>
1. Identify the given values:
   - Rate of electron transfer = 10^10 electrons/second
   - Total charge to transfer Q = 1 C
   - Elementary charge e = 1.6 * 10^-19 C/electron
2. Calculate charge transferred per second:
   - dq/dt = (10^10 electrons/s) * (1.6 * 10^-19 C/electron) = 1.6 * 10^-9 C/s (1.6 nC/s)
3. Calculate time t required:
   - t = Q / (dq/dt) = 1 C / (1.6 * 10^-9 C/s) = 6.25 * 10^8 seconds
4. Convert seconds to years:
   - 1 year ≈ 3.154 * 10^7 seconds
   - t (years) = (6.25 * 10^8) / (3.154 * 10^7) ≈ 19.8 years
</think>

Answer: The charge transferred per second is 1.6 × 10⁻⁹ C/s. The total time required is 6.25 × 10⁸ seconds, which is approximately 19.8 years.

Explanation: Electrons carry a fundamental unit of charge equal to 1.6 × 10⁻¹⁹ C. With a transfer rate of 10¹⁰ electrons per second, the total charge moving per second is 1.6 × 10⁻⁹ C/s. To accumulate 1 Coulomb of charge, dividing the total required charge by the transfer rate yields 6.25 × 10⁸ seconds (~19.8 years).

🔬 Frameworks & Tools

Unsloth: 4-bit QLoRA fine-tuning and inference.
PEFT: LoRA adapter implementation.
TRL: SFTTrainer for Supervised Fine-Tuning.
Hugging Face Transformers: Tokenizer and causal LM architecture.
Weights & Biases: Run monitoring and experiment logs.

text

<|im_start|>system
Follow these steps to answer the user's question:
1. First, reason step-by-step inside <think> tags. This is your internal monologue.
2. After the </think> tag, provide a complete, helpful, and polished answer for the user that includes both a direct answer and an explanation.<|im_end|>
<|im_start|>user
{Question}<|im_end|>
<|im_start|>assistant
<think>
{Step-by-step reasoning and physics derivations}
</think>

Answer: {Direct Answer}
Explanation: {Detailed Explanation}<|im_end|>

Step

Training Loss

Validation Loss

0.8160

0.5512

0.5019

0.4574

0.4581

0.4241

0.4187

0.4070

0.4173

0.3952

0.4043

python

from unsloth import FastLanguageModel
from transformers import TextStreamer

# Load the base model with adapter weights
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "Rakesh7n/Qwen3_4B_NCRT_Physics_12th_Finetuned", # points to this LoRA adapter repo
    max_seq_length = 2048,
    load_in_4bit = True,
)
FastLanguageModel.for_inference(model)

# Prepare prompt
messages = [
    {
        "role": "user",
        "content": "If 10^10 electrons move from one body to another every second, how long will it take to transfer 1 C of charge?"
    }
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize = False,
    add_generation_prompt = True,
)

inputs = tokenizer(text, return_tensors = "pt").to("cuda")
streamer = TextStreamer(tokenizer, skip_prompt = False)

# Generate with reasoning
_ = model.generate(
    **inputs,
    max_new_tokens = 2048,
    temperature = 0.7,
    top_p = 0.8,
    top_k = 20,
    streamer = streamer,
)

python

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

base_model_id = "unsloth/Qwen3-4B-Thinking-2507"
adapter_model_id = "Rakesh7n/Qwen3_4B_NCRT_Physics_12th_Finetuned"

tokenizer = AutoTokenizer.from_pretrained(base_model_id)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.float16,
    device_map="auto"
)

# Load PEFT LoRA adapter on top of base model
model = PeftModel.from_pretrained(base_model, adapter_model_id)

messages = [
    {
        "role": "system",
        "content": "Follow these steps to answer the user's question:\n1. First, reason step-by-step inside <think> tags. This is your internal monologue.\n2. After the </think> tag, provide a complete, helpful, and polished answer for the user that includes both a direct answer and an explanation."
    },
    {
        "role": "user",
        "content": "What is the formula for electric flux according to Gauss's Law?"
    }
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

outputs = model.generate(
    **inputs,
    max_new_tokens=1024,
    temperature=0.7,
    top_p=0.8,
    do_sample=True
)

print(tokenizer.decode(outputs[0], skip_special_tokens=False))

text

<think>
1. Identify the given values:
   - Rate of electron transfer = 10^10 electrons/second
   - Total charge to transfer Q = 1 C
   - Elementary charge e = 1.6 * 10^-19 C/electron
2. Calculate charge transferred per second:
   - dq/dt = (10^10 electrons/s) * (1.6 * 10^-19 C/electron) = 1.6 * 10^-9 C/s (1.6 nC/s)
3. Calculate time t required:
   - t = Q / (dq/dt) = 1 C / (1.6 * 10^-9 C/s) = 6.25 * 10^8 seconds
4. Convert seconds to years:
   - 1 year ≈ 3.154 * 10^7 seconds
   - t (years) = (6.25 * 10^8) / (3.154 * 10^7) ≈ 19.8 years
</think>

Answer: The charge transferred per second is 1.6 × 10⁻⁹ C/s. The total time required is 6.25 × 10⁸ seconds, which is approximately 19.8 years.

Explanation: Electrons carry a fundamental unit of charge equal to 1.6 × 10⁻¹⁹ C. With a transfer rate of 10¹⁰ electrons per second, the total charge moving per second is 1.6 × 10⁻⁹ C/s. To accumulate 1 Coulomb of charge, dividing the total required charge by the transfer rate yields 6.25 × 10⁸ seconds (~19.8 years).

Qwen3_4B_NCRT_Physics_12th_Finetuned

README

📌 Model Overview

🎯 Intended Use & Limitations

Primary Use Cases

Limitations & Out-of-Scope

📊 Dataset Details

Conversation Schema

📈 Training & Validation Loss

🚀 How to Run Inference

1. Using Unsloth (Loading Base Model + Adapter)

2. Using Standard Hugging Face `transformers` & `peft`

📝 Example Output

User Prompt:

Model Output:

🔬 Frameworks & Tools

Explore FriendliAI today

README

📌 Model Overview

🎯 Intended Use & Limitations

Primary Use Cases

Limitations & Out-of-Scope

📊 Dataset Details

Conversation Schema

📈 Training & Validation Loss

🚀 How to Run Inference

1. Using Unsloth (Loading Base Model + Adapter)

2. Using Standard Hugging Face `transformers` & `peft`

📝 Example Output

User Prompt:

Model Output:

🔬 Frameworks & Tools

Qwen3_4B_NCRT_Physics_12th_Finetuned

README

📌 Model Overview

🎯 Intended Use & Limitations

Primary Use Cases

Limitations & Out-of-Scope

📊 Dataset Details

Conversation Schema

📈 Training & Validation Loss

🚀 How to Run Inference

1. Using Unsloth (Loading Base Model + Adapter)

2. Using Standard Hugging Face transformers & peft

📝 Example Output

User Prompt:

Model Output:

🔬 Frameworks & Tools

Explore FriendliAI today

README

📌 Model Overview

🎯 Intended Use & Limitations

Primary Use Cases

Limitations & Out-of-Scope

📊 Dataset Details

Conversation Schema

📈 Training & Validation Loss

🚀 How to Run Inference

1. Using Unsloth (Loading Base Model + Adapter)

2. Using Standard Hugging Face transformers & peft

📝 Example Output

User Prompt:

Model Output:

🔬 Frameworks & Tools

2. Using Standard Hugging Face `transformers` & `peft`

2. Using Standard Hugging Face `transformers` & `peft`