tinyllama-lora-fast

<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->

tinyllama-lora-fast

Trained primarily on Prothom Alo news data, this model naturally writes in that newspaper’s concise, reportorial Bangla style.

This model is a fine-tuned version of TinyLlama/TinyLlama-1.1B-Chat-v1.0 Bangla Newspaper Dataset. The single goal: enable fast training and low VRAM usage—even on small GPUs or TPUs.Use Dynamic Padding + Token Budget for low Gpu And Tpu It achieves the following results on the evaluation set:

• Loss: 0.8744

Model description

Base model: TinyLlama/TinyLlama-1.1B-Chat-v1.0

Adapter: PEFT-LoRA (r=16, alpha=32, dropout=0.05, bias=none)

Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj

Context length (train max): MAX_LEN=768

Optimization: AdamW (lr≈3e-5, warmup≈0.03, weight_decay≈0.05)

Batching: length-based bucketing + dynamic padding

Precision: fp16 inference-ready (training setup Kaggle/Colab-friendly)

Decoding (low hallucination preset): temperature=0.0, do_sample=False, no_repeat_ngram_size=4, repetition_penalty≈1.1–1.15

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 3e-05
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 4
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.03
• num_epochs: 2
• mixed_precision_training: Native AMP

Training results

Framework versions

• PEFT 0.12.0
• Transformers 4.45.2
• Pytorch 2.6.0+cu124
• Datasets 4.1.1
• Tokenizers 0.20.3