You can deploy this model and start earning money today!

AsifZaman1912/
ChatgaiyyaLM-0.5B-B2_subset_ctg_syl_bn

Name: ChatgaiyyaLM-0.5B-B2_subset_ctg_syl_bn
Availability: InStock

Views

No views yet

ChatgaiyyaLM — B2_subset_ctg_syl_bn

LoRA adapter adapting Qwen/Qwen2.5-0.5B to Chittagonian (চাটগাঁইয়া), a Bangla dialect with tens of millions of speakers and almost no NLP resources.

One arm of a controlled ablation. Pipeline, benchmark and full results: https://github.com/A-K-M-Asifuzzaman/Chatgaiya-AI

Results

chrF on the gold translation track of ChatgaiyyaBench, 120 fixed-seed items per direction.

System	ctg → bn	bn → ctg
base (no adaptation)	20.54	19.77
A1_vocab_only	56.69	49.26
A2_subset_cpt	59.62	51.04
A3_full_cpt	59.33	51.98
B1_subset_ctg_syl	62.89	50.00
B2_subset_ctg_syl_bn	65.64	48.64
transducer (rules+lexicon, no GPU)	—	57.94
copy input unchanged	29.51	—

Read the transducer row. A mined rule system costing no GPU time beats the neural arms at Bangla → Chittagonian. That is a real result about the difficulty of the task, reported rather than omitted.

An unadapted model is worse than doing nothing: base scores 20.54 on ctg → bn; copying the input scores 29.51.

This arm


Adaptation	`vocab+subset_cpt`
Dialects in CPT	ctg, syl, bn
CPT tokens	8,017,920
SFT loss	1.1264
Wall time	117 min
Added vocabulary	3,914 Chittagonian subword tokens
Precision	fp16+GradScaler (Tesla T4, sm_75 has no hardware bf16)

Usage

python

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

REPO = "AsifZaman1912/ChatgaiyyaLM-0.5B-B2_subset_ctg_syl_bn"
tok = AutoTokenizer.from_pretrained(REPO)
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-0.5B", dtype=torch.float16)
model.resize_token_embeddings(len(tok))     # the adapter adds 3,914 tokens
model = PeftModel.from_pretrained(model, REPO).eval()

prompt = "চাটগাঁইয়া থেকে প্রমিত বাংলায় অনুবাদ করো:\nঅ্যাঁই ভাত ন হাইয়্যুম\nঅনুবাদ:"
ids = tok(prompt, return_tensors="pt")
out = model.generate(**ids, max_new_tokens=64, do_sample=False)
print(tok.decode(out[0][ids["input_ids"].shape[1]:], skip_special_tokens=True))

resize_token_embeddings before loading is required — the adapter carries a larger embedding matrix than the base model. Always load the tokenizer from this repo: token ids were assigned in insertion order, so a rebuilt tokenizer silently misaligns every added token.

Limitations

0.5B parameters. Qwen2.5-1.5B needs ~14.5 GB before activations under PEFT's duplicated embeddings and OOMs a 16 GB T4.
Fixed instruction-tuning budget, not full epochs.
Four of eight benchmark tracks are silver (auto-derived labels or transduced text), awaiting native-speaker validation. Do not read absolute numbers off them.
Research artifact, not a production translator.