makireddyvighnesh

Qwen2.5-1.5B-Instruct-telugu-translit-grpo

README

License: apache-2.0

Results

Aksharantar Telugu test, all 10,260 words, greedy decoding. Reward is the environment's: 0.8 × (1 − CER) + 0.2 × exact match, and 0 for an empty or looping answer.

Table with columns: base, this adapter
	base	this adapter
reward	0.372	0.695
character error rate (capped at 1 per word)	0.547	0.204
exact match	4.6%	29.1%
chrF	0.269	0.527

By where the word comes from (reward / exact match):

Table with columns: origin, words, base, this adapter
origin	words	base	this adapter
Dakshina (common words)	5,724	0.386 / 6.0%	0.721 / 32.9%
AK-Freq (frequent words)	2,241	0.413 / 3.8%	0.776 / 37.1%
AK-NEI (Indian named entities)	1,347	0.334 / 2.5%	0.566 / 13.7%
AK-NEF (foreign named entities)	948	0.243 / 1.3%	0.526 / 8.9%

Sampling 8 answers at temperature 0.9 on 1,000 test words, the best of 8 reaches 0.737 (base 0.509), and at least one answer is exactly right for 35% of words (base 8%).

The base numbers are the untuned model on the same words, scored the same way. They are aggregates, not a word-by-word pairing.

What it does and doesn't get right

Twelve test words, three per origin, drawn at random (seed 1), not picked by hand:

Table with columns: origin, romanized, reference, this adapter, CER
origin	romanized	reference	this adapter	CER
Dakshina	saalid	సాలిడ్	సాలిడ్	0.000
Dakshina	chendinade	చెందినదే	చెందినడే	0.125
Dakshina	thevadam	తేవడం	తీవడం	0.200
AK-Freq	gurtunchukovadam	గుర్తుంచుకోవడం	గుర్తుంచుకోవడం	0.000
AK-Freq	padutunnaayi	పడుతున్నాయి	పదుతున్నాయి	0.091
AK-Freq	sariayina	సరిఅయిన	సరియిన	0.143
AK-NEI	nerolac	నేరోలాక్	నేరోలచ్	0.250
AK-NEI	nischal	నిశ్చల్	నిశ్ల	0.286
AK-NEI	fatima	ఫాతిమా	ఫతిమ	0.333
AK-NEF	kenner	కెన్నర్	కేన్నర్	0.143
AK-NEF	smith	స్మిత్	స్మింగ్	0.333
AK-NEF	greenville	గ్రీన్విల్లే	గ్రీవలన్	0.500

Most remaining errors are the ones romanization itself makes ambiguous: vowel length (e/ee, a/aa) and retroflex versus dental consonants (d → డ or ద). Named entities, foreign ones most of all, stay hardest, because their spelling in Telugu follows pronunciation rather than the Latin letters. The model sometimes ends a word with a zero-width non-joiner; scoring removes it, as Telugu text normalization usually does.

Use it

python

import torchfrom peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer base = "Qwen/Qwen2.5-1.5B-Instruct"tokenizer = AutoTokenizer.from_pretrained(base)model = AutoModelForCausalLM.from_pretrained(base, dtype=torch.float16, device_map="auto")model = PeftModel.from_pretrained(model, "makireddyvighnesh/Qwen2.5-1.5B-Instruct-telugu-translit-grpo") prompt = ("Transliterate the following romanized Telugu into Telugu script. "          "Output only the Telugu text.\n\nnakshatram")inputs = tokenizer.apply_chat_template([{"role": "user", "content": prompt}],                                       add_generation_prompt=True, return_tensors="pt",                                       return_dict=True).to(model.device)output = model.generate(**inputs, max_new_tokens=48, do_sample=False)print(tokenizer.decode(output[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Use the prompt exactly as above: it is the one the model was trained on. The adapter was trained on single words; for a sentence, transliterate it word by word.

With vLLM:

bash

vllm serve Qwen/Qwen2.5-1.5B-Instruct --enable-lora --max-lora-rank 16 \  --lora-modules telugu-translit=makireddyvighnesh/Qwen2.5-1.5B-Instruct-telugu-translit-grpo

How it was trained

Environment. An OpenEnv server that serves one transliteration task per episode: reset returns a romanized word and the prompt, step grades one answer and ends the episode. It holds the references; the trainer only ever sees task IDs. The task set is a snapshot of Aksharantar Telugu pinned to dataset revision e418c1f (snapshot 0936c463…, recorded in training/environment.json), keeping only the human-romanized sources:

Table with columns: split, words
split	words
train	50,927 (Dakshina)
validation	3,716 (Dakshina)
test	10,260 (Dakshina, AK-Freq, AK-NEI, AK-NEF)

Aksharantar's Telugu train split is otherwise 95% IndicCorp, web-mined and machine-romanized, with about 1 in 8 native words misspelled in a 40-row audit. Those 2.38M pairs were left out.

Reward. 0.8 × (1 − CER) + 0.2 × exact match. CER is the edit distance over Unicode code points after normalization (NFKC/NFC, zero-width characters and quotes removed, Telugu digits to ASCII), divided by the reference length and capped at 1. A word-level rate would score one wrong vowel sign as a total miss. Empty answers and answers that loop or balloon score 0, so the policy can't farm partial credit by repeating itself. <think> blocks and a wrapping code fence are stripped; prose around the answer is not, so the policy learns to output only the word.

GRPO. TRL 1.16 GRPOTrainer with environment_factory: one environment client per rollout, reset(task_id) before generation, step(completion) in the reward function.

Table with columns: setting, value
setting	value
steps	1000 (about 1 epoch over 16,000 train words)
prompts × answers per step	16 × 8
learning rate	5e-5, cosine, 50 warmup steps
LoRA	r 16, alpha 32, dropout 0, all linear layers
temperature	0.7
max completion length	48 tokens (longest reference: 37)
loss	DAPO (TRL default), no KL term
precision	fp16 on 2 × NVIDIA T4 (Kaggle)
wall time	2.8 h of training

Validation reward (256 frozen validation words, 8 samples each) rose from 0.47 at step 50 to 0.70 by step 650, then stayed flat. By the end, 58% of groups gave no learning signal (all 8 answers scored the same) and policy entropy had fallen from 0.33 to 0.06, so more steps on this data would add little.

A second run that called the same reward function in-process instead of through the environment, with the same seed and words, produced the same test scores to four decimals. That is the check that the environment serves and grades exactly what the reward function computes.

Files

adapter_model.safetensors, adapter_config.json: the LoRA adapter
training/log_history.json: every logged metric, every step
training/test_report.json: the test evaluation above
training/environment.json: the environment snapshot and Aksharantar revision trained against
assets/curves.png: the figure at the top

Limitations

Single words only. It was never trained on phrases or sentences.
Romanization varies from person to person. The test references are one spelling per word, so a valid alternative spelling counts as an error.
Named entities, especially foreign ones, are much weaker than common words.
Evaluated on Aksharantar only.

License and credit

The adapter is Apache-2.0, like the base model. Training and test words come from Aksharantar by AI4Bharat, CC BY 4.0. The environment's design follows the FineEnvs multilingual OCR and ASR environments.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

makireddyvighnesh

Model Tree

Base

Qwen/Qwen2.5-1.5B-Instruct

Adapter

this model

Input Modalities

Text

Output Modalities