Parameter	Value
LoRA rank	32
LoRA alpha	32
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Trainable params	~9.7M / 135M (6.77%)
Quantization	4-bit (QLoRA via Unsloth)
Batch size	32
Gradient accumulation	2 (effective batch: 64)
Learning rate	2e-4 (linear decay)
Warmup steps	100
Epochs	10
Sequence length	512 tokens
Chunking	256-word chunks, 20% overlap, packed
Hardware	NVIDIA RTX 4090
Training time	~14 min

Training Data

188 arXiv ML papers (2024–2026), downloaded via the arXiv API
Papers cleaned: references section removed, appendix preserved
Split: 138 train / 50 validation
After chunking + packing: ~5,200 training sequences

Evaluation Results

Evaluated on 50 held-out papers, 50 samples, 20-word prefix → 50-word generation:

Metric	Base Model	This Model	Δ
Perplexity	22.97	18.36	-20.1%
Cross-Entropy	3.134	2.910	-7.1%
ROUGE-1	0.178	0.213	+19.7%
ROUGE-L	0.114	0.143	+25.4%
BERTScore F1	0.736	0.753	+2.3%
BLEU	0.016	0.022	+37.5%

Key Findings from Experiment Loop

This model was selected as the winner from a systematic experiment loop:

LoRA beats full fine-tuning on small datasets — 138 papers is too few for full FT; LoRA's regularisation helps
Rank doesn't matter much — r=8/16/32 all plateau at the same eval loss; data is the bottleneck
Interleaving with large HF datasets didn't help at this data scale — domain signal gets diluted

How to Use

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model_id = "HuggingFaceTB/SmolLM-135M"
adapter_id = "JaydeepR/SmolLM-135M-CPT-LoRA-r32"

tokenizer = AutoTokenizer.from_pretrained(base_model_id)
model = AutoModelForCausalLM.from_pretrained(base_model_id)
model = PeftModel.from_pretrained(model, adapter_id)

prompt = "We propose a novel attention mechanism that"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=100, repetition_penalty=1.2)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Limitations

Trained on only 188 papers — sufficient for stylistic adaptation, not factual knowledge
May hallucinate scientific content (model learns paper structure, not paper facts)
Optimised for ML paper generation; may not generalise to other scientific domains
135M parameter model — limited overall capability

Citation

@misc{smollm135m-cpt-lora,
  author = {Jaydeep Raijada},
  title  = {SmolLM-135M CPT LoRA r=32 — Continued Pre-Training on arXiv ML Papers},
  year   = {2026},
  url    = {https://huggingface.co/JaydeepR/SmolLM-135M-CPT-LoRA-r32}
}