Telugu is spoken by over 80 million people but remains severely underrepresented in modern LLM training data. Most large language models produce broken grammar, mixed-script outputs, or refuse to respond in Telugu entirely. This creates a digital divide where Telugu speakers cannot access AI-powered tools in their native language.
This model solves that by adapting a 120B-parameter model to achieve robust Telugu language competence across diverse tasks — from news summarization and QA to creative writing and translation — demonstrating that targeted, high-quality data adaptation via AutoScientist can transform a primarily English LLM into a capable Telugu-speaking assistant.
The model was adapted using the data-centric tools on the AutoScientist platform:
Adaptive Data Pipeline: Cleaned Telugu text encoding, removed broken Unicode, and normalized script inconsistencies. This automatic data cleaning upgraded the quality score from Grade C (6.0) to Grade B (8.14), a +35.6% relative improvement.
Reasoning Traces: Guided the model to output logical chain-of-thought steps in Telugu before presenting final summaries, translations, or answers.
Hyperparameter Optimization: Orchestrated optimal LoRA hyperparameter search and managed the training loop for the 120B parameter base model.
Evaluation: Evaluated our adapted model against the base model on unseen Telugu prompts, with the judge scoring based on the quality of language competence.
Overall License: CC-BY-SA-3.0 (respecting the most restrictive source)
Adaptive Quality Evaluation:
Metric
Before
After
Change
Grade
C
B
⬆️
Score
6.0
8.14
+35.6%
Percentile
—
18.23
—
⚙️ Training Configuration
Hyperparameter
Value
Finetune Job ID
4b0833cd-fbba-4d94-85a3-77945a254cea
Training Experiment ID
a9910567-141b-4b1a-ad6e-8216e18dc6bd
LoRA Rank (r)
64
LoRA Alpha
128
LoRA Dropout
0
Target Modules
q_proj, k_proj, v_proj, o_proj
Epochs
3
Batch Size
max
Learning Rate
3e-4
LR Scheduler
Cosine (0.5 cycles)
Warmup Ratio
0.05
Weight Decay
0.01
Max Grad Norm
1.0
Min LR Ratio
0.1
📈 Evaluation Results
LLM-as-a-Judge Win Rate:66% adapted vs. 34% base model
The adapted model demonstrates significantly improved Telugu language competence, outperforming the base model on unseen Telugu prompts judged on grammar, vocabulary, and natural language quality.
Intended Use Cases
Telugu news headline and article generation
Question answering based on Telugu news and Andhra Pradesh facts
Creative Telugu poetry completion
English to Telugu translation
Telugu instruction following and general assistance
