Molly Specialist — Computer Science NLP

Answers technical NLP questions with more accurate explanations of transformer architectures, tokenization algorithms, and evaluation metrics than the base model.

Part of Molly, an orchestrator that keeps a library of small domain specialists over one quantized base and routes each request to the right one, so a single machine answers across many fields without loading a separate large model for each.

What this specialist handles well

Explaining attention mechanisms and transformer architecture internals accurately
Comparing tokenization algorithms like BPE, WordPiece, and SentencePiece
Clarifying NLP evaluation metrics such as BLEU, ROUGE, and BERTScore

Try it with

"How does rotary position embedding differ from absolute positional encoding in transformers?"
"What are the trade-offs between byte-pair encoding and unigram tokenization for multilingual models?"
"How do I implement contrastive loss for fine-tuning a sentence embedding model?"

Before you run: the base model is gated

This adapter needs the base weights, and the base is access-gated. Do this once:

Accept the base licence: https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct
Create a read token: https://huggingface.co/settings/tokens
Make the token available:
- Google Colab: Secrets panel (key icon) → Add new secret → name HF_TOKEN, enable Notebook access.
- Kaggle: Add-ons → Secrets → add HF_TOKEN.
- Local: huggingface-cli login or export HF_TOKEN=...

Skipping this gives GatedRepoError / 401 Unauthorized when the base loads. A stored Colab secret is not applied automatically — authenticate in code, as below.

Quickstart

# pip install -U transformers peft accelerate
import os, torch
from huggingface_hub import login
try:
    from google.colab import userdata
    login(userdata.get("HF_TOKEN"))
except Exception:
    tok = os.environ.get("HF_TOKEN")
    login(tok) if tok else login()

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

BASE = "meta-llama/Llama-3.1-8B-Instruct"
ADAPTER = "BoomJules/molly-cs-nlp"

tok = AutoTokenizer.from_pretrained(BASE)
base = AutoModelForCausalLM.from_pretrained(BASE, torch_dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(base, ADAPTER).eval()

msgs = [{"role": "user", "content": "Your question here"}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(ids, max_new_tokens=300)
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))

Low-VRAM (4-bit) — fits a free Colab/Kaggle GPU (~6–7 GB)

# pip install -U transformers peft accelerate bitsandbytes
import os, torch
from huggingface_hub import login
try:
    from google.colab import userdata
    login(userdata.get("HF_TOKEN"))
except Exception:
    login(os.environ.get("HF_TOKEN"))

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

bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4",
                         bnb_4bit_compute_dtype=torch.bfloat16, bnb_4bit_use_double_quant=True)
tok = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")
base = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.1-8B-Instruct", quantization_config=bnb, device_map="auto")
model = PeftModel.from_pretrained(base, "BoomJules/molly-cs-nlp").eval()

Adapter details


Base model	`meta-llama/Llama-3.1-8B-Instruct`
Method	LoRA (PEFT)
Rank / alpha	32 / 64
Domain	Computer Science NLP

Troubleshooting

GatedRepoError / 401 Unauthorized — base licence not accepted, or HF_TOKEN missing, or the Colab secret was stored but login(...) was never called.
CUDA out of memory — use the 4-bit snippet on a GPU runtime.
Adapter seems to have no effect — confirm the base id matches base_model above.