uditjain/lori-qwen2.5-1.5b-science

LoRI Science Adapter for Qwen2.5-1.5B-Instruct

Model Description

This repository contains a PEFT LoRA adapter trained by uditjain as part of the mewtwo research line on low-rank expert factorization and multi-domain adapter composition.

The adapter was trained on top of Qwen/Qwen2.5-1.5B-Instruct using a custom LoRI-style procedure:

• inject standard LoRA modules into the base model
• replace the LoRA down-projection with a shared frozen random projection
• keep the corresponding up-projection trainable as the domain-specific factor
• apply post-training DARE-style sparsification

This release is the domain adapter artifact only. It is not a standalone model.

What Is In This Repo

• adapter_model.safetensors: adapter weights
• adapter_config.json: PEFT adapter configuration
• artifacts/training_state.json: saved training configuration/state snapshot
• artifacts/training_log.json: summarized training log

Base Model

• Base model: Qwen/Qwen2.5-1.5B-Instruct
• Base model license: apache-2.0
• PEFT type: LORA
• Rank: 32
• Alpha: 64
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj

Intended Use

Science tutoring, factual explanation, and principle-based question answering on top of Qwen2.5-1.5B-Instruct.

Training Data

This adapter was trained from the prepared domain corpus in mewtwo/data/lori_moe.

• Final prepared example count: 11679
• Average tokenized sequence length statistic: 742.2
• Source datasets / preparation path:
  • allenai/sciq

The training pipeline that prepared these datasets is documented in the mewtwo repository under src/lori_moe/data/prepare_datasets.py.

Training Procedure

• Training method: LoRI-style PEFT LoRA adaptation with shared frozen random projection and trainable domain-specific factor
• Post-processing: DARE-style sparsification
• Precision: BF16 mixed-precision training
• Optimizer: bnb_paged_adamw_8bit
• Best recorded training loss: 1.359215
• Recorded training time: 22.59 minutes

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model_id = "Qwen/Qwen2.5-1.5B-Instruct"
adapter_id = "uditjain/lori-qwen2.5-1.5b-science"

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

Research Context

This adapter belongs to a 5-domain family of LoRI-trained experts (math, code, science, legal, medical) explored in the mewtwo project as a successor direction to the earlier Synapta prompt-level composition work.

One family-level empirical result from the saved adapter weights is very low cross-domain overlap for the final Qwen/Qwen2.5-1.5B-Instruct expert set:

• average absolute cross-domain cosine similarity was measured at approximately 0.00685 across the 5 published Qwen2.5-1.5B expert adapters.

This supports the geometric motivation for the method, but it does not by itself prove superior end-to-end reasoning performance.

Limitations

• This release is a domain-steering adapter, not a full validated routed MoE system.
• The surrounding router and end-to-end composition stack remain a separate research layer.
• Domain specialization does not guarantee factual correctness.
• Scientific explanations may omit caveats or oversimplify domain nuances and should not be treated as authoritative without checking sources.
• Legal and medical use cases require especially careful human oversight.

Open Source Notes

This release is intended to make the trained adapter artifact reproducible and reusable by others working on:

• low-rank expert factorization
• adapter composition
• domain adaptation on small language models
• PEFT-based research baselines

If you build on this work, please cite the future paper or link back to the original mewtwo research repository when available.