agentic-rag-lora

LoRA adapter only (not full base weights) for HuggingFaceTB/SmolLM-135M, lightly trained on CPU with short instruction/response texts about agentic systems and RAG.

Comparison note: This is a tiny PEFT adapter (~SmolLM-135M base), not a 4-bit 7B chat model and not a paid Gradio Space. Load the base model, then attach this adapter.

Important

• Repository contains adapter weights + tokenizer config, not a full 135M rehost of new authorship.
• You must load the base SmolLM-135M and then attach this adapter.
• Intended as a small demo of domain-adapted instruction style for agentic/RAG docs — not a production LLM.

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "HuggingFaceTB/SmolLM-135M"
tok = AutoTokenizer.from_pretrained("hharsha/agentic-rag-lora")
model = AutoModelForCausalLM.from_pretrained(base)
model = PeftModel.from_pretrained(model, "hharsha/agentic-rag-lora")

prompt = "Explain hybrid search in a RAG pipeline:"
inputs = tok(prompt, return_tensors="pt")
print(tok.decode(model.generate(**inputs, max_new_tokens=64)[0], skip_special_tokens=True))

Training summary

Base	HuggingFaceTB/SmolLM-135M
Method	PEFT LoRA (r=8, alpha=16, q_proj/v_proj)
Epochs	1 (CPU, float32)
Trainable params	~461k (0.34%)
Data	Showcase project summaries + synthetic agentic/RAG instructions

Related

• Embeddings: hharsha/agentic-systems-minilm
• Tag generator: hharsha/agentic-github-tagger
• Dataset: hharsha/agentic-systems-showcase
• Meta tags data: hharsha/agentic-github-meta
• Studio: https://agentic-systems-studio.com