You need to agree to share your contact information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
By requesting access you agree to the CC BY-NC 4.0 license terms and acknowledge the Meta LLaMA 2 License for the base model. Commercial use requires a separate licence from IIT Delhi.
Log in or Sign Up to review the conditions and access this model content.
RSL-SETU-LoRA-v35 — IKS Teaching Language Model (LoRA Adapter)
SETU (सेतु — Bridge, as in Rāma Setu) bridges Indian Knowledge System pedagogical techniques and modern education through fine-tuned language models.
RSL-SETU-LoRA-v35 is a QLoRA adapter for ai4bharat/Airavata (7B, LLaMA-based) fine-tuned on 13,882 IKS instruction pairs to teach using 40+ traditional Indian pedagogical techniques.
Model Summary
| Property | Value |
|---|---|
| Base model | ai4bharat/Airavata (7B params, LLaMA architecture) |
| Adapter type | QLoRA (4-bit NormalFloat quantization) |
| Adapter size | ~64 MB (safetensors) |
| SHA-256 | 7f6cc4e26e905fa30f9131108992a267fe7da71d4363a189886755dd110485aa |
| Trainable params | < 1% of base model |
| LoRA rank | 16 |
| LoRA alpha | 32 |
| Target modules | q_proj, v_proj, k_proj, o_proj |
| Training data | 13,882 IKS instruction pairs (v1.1 dataset) |
| Epochs | 3 |
| Max sequence length | 512 tokens |
| License | CC BY-NC 4.0 |
What This Model Does
Given a student scenario with cognitive state metadata, RSL-SETU-LoRA-v35 generates pedagogically appropriate responses using IKS techniques. It can:
- Apply the correct IKS technique for a given cognitive state and topic
- Explain technique selection — why one technique works and another doesn't
- Handle multi-turn dialogues with cognitive load progression
- Teach across subjects — Math, Science, Social Studies, Literature (CBSE 6–12 + UG)
- Respond in multiple languages — English, Hindi, Sanskrit, Tamil, Telugu, Kannada, Malayalam
Example
Input:
A Class 9 student with high cognitive load (0.9) and low attention (0.3) is learning number systems in Mathematics. Apply the appropriate IKS technique.
Output:
[Applying Dhyana-based Focus Protocol] Let's pause and center your attention before we continue. Take a moment to focus... Now, number systems are like the building blocks of mathematics. In the tradition of Vedic mathematics, numbers were understood through patterns...
Evaluation Results
RSL-PRAJNA v2 Benchmark (351 questions, judge: llama-3.3-70b)
| Dimension | Score |
|---|---|
| Overall | 7.35 |
| Technique Fidelity | 6.64 |
| Pedagogical Quality | 6.44 |
| Multilingual Quality | 6.31 |
| Safety | 9.99 |
Rank 4 of 19 on the v2 leaderboard — outperforms llama-3.1-70b (7.28), llama-3.3-70b (7.27), and Qwen-IKS-v1 (6.82). Highest multilingual quality (6.31) among all LoRA variants.
Training Data Composition (v1.1 Dataset)
| Source | Count | Description |
|---|---|---|
| v1.0 base | 9,890 | IKS instruction pairs (40+ techniques) |
| Bhagavad Gita | 2,100 | 700 shlokas × 3 pedagogical formats |
| Sangam literature | 1,892 | Tamil Sangam corpus IKS pairs |
| Miscellaneous enrichment | ~300 | Upanishads, Arthashastra, edge cases |
| Total | 13,882 |
Training Configuration
| Parameter | Value |
|---|---|
| Quantization | 4-bit NF4 (double quantization) |
| Optimizer | paged_adamw_32bit |
| Learning rate | 2e-4 (cosine schedule) |
| Warmup ratio | 0.03 |
| Gradient accumulation | 16 steps |
| Gradient checkpointing | Enabled |
| Hardware | NVIDIA L4 GPU (Vertex AI) |
How to Use
With PEFT + Transformers
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base_model = AutoModelForCausalLM.from_pretrained(
"ai4bharat/Airavata",
load_in_4bit=True,
device_map="auto"
)
model = PeftModel.from_pretrained(base_model, "RSL-INTRINSICLab-IIT/RSL-SETU-LoRA-v35")
tokenizer = AutoTokenizer.from_pretrained("ai4bharat/Airavata")
prompt = """Below is an instruction that describes a task. Write a response.
### Instruction:
A Class 10 student is learning quadratic equations. Their cognitive load is moderate (0.5).
Apply Nikhilam Sutra (Vedic Mathematics) to explain solving x² - 7x + 12 = 0.
### Response:"""
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Limitations
- Domain-specific — Optimized for IKS pedagogy, may underperform on general tasks
- Latency — 40–48 seconds per response on L4 GPU (7B model + LoRA merge overhead)
- Multilingual — Best in English and Hindi; Tamil/Sanskrit/Telugu responses are functional but less fluent
- Context window — 512 tokens max (inherited from training configuration)
- Prompt injection — The model does not include built-in defenses against prompt injection or adversarial inputs. In production deployments, wrap the model behind an input-sanitization layer that filters system-prompt overrides, role-play injection attempts, and encoded payloads before they reach the model. Do not expose the raw model endpoint to untrusted user input without guardrails.
Citation
@model{rsl_setu_lora_v35,
title={RSL-SETU-LoRA-v35: QLoRA Adapter for IKS Teaching Language Model},
author={Sivasubramani, Santhosh},
year={2026},
institution={INTRINSIC Lab, RSL Foundation, IIT Delhi},
base_model={ai4bharat/Airavata},
url={https://huggingface.co/RSL-INTRINSICLab-IIT/RSL-SETU-LoRA-v35}
}
Related Resources
- RSL-PRAJNA-v2 — Evaluation benchmark used to score this model
- RSL-SETU-Classifier-15M — Lightweight IKS technique classifier (95.6% accuracy)
- RSL-BHARATI-v3 — Multilingual tokenizer (7 languages, 32K vocab)
- RSL-SHRUTI-Thirukkural — Thirukkural-CBSE mapping dataset
Patent Notice
The adaptive learning orchestration method implemented in this model is covered by Indian Complete Patent Application No. 1536IN243 — "Systems and Methods for Real-Time Adaptive Learning Orchestration", filed March 2026 in the name of Indian Institute of Technology Delhi. Inventor: Prof. Santhosh Sivasubramani.
The model weights are released under CC BY-NC 4.0 for research and educational use. The patented method may not be used in commercial products or services without a separate licence from IIT Delhi.
License
Adapter weights: CC BY-NC 4.0 — Free for research and educational use. Commercial use requires a license from IIT Delhi.
Base model: This adapter requires ai4bharat/Airavata, which is distributed under the Meta LLaMA 2 License. Users must independently accept the LLaMA 2 terms on the Airavata model page before use. The combined use of this adapter with the base model is subject to both licenses.
Acknowledgment
Demonstrated at the Bharat Bodhan AI Conclave, anchored and driven by the Ministry of Education and IIT Madras, New Delhi.
Contact
Prof. Santhosh Sivasubramani Director, INTRINSIC Laboratory RSL Foundation, Centre for SeNSE, IIT Delhi ssivasub@iitd.ac.in https://intrinsic.iitd.ac.in
- Downloads last month
- 1
Model tree for RSL-INTRINSICLab-IIT/RSL-SETU-LoRA-v35
Base model
ai4bharat/Airavata