adaption-market-analysis-final-model

Adaption Market Analysis Final Model

This repository contains the PEFT LoRA adapter produced in Attempt 2 of our AutoScientist Challenge project. The model was adapted for analyst-style finance responses using the 45,758-row Final dataset created with Adaption Labs' Adaptive Data and Combine workflows.

Artifact

• Base model: togethercomputer/Llama-4-Scout-17B-16E-Instruct_bnb_4bit
• Adapter type: LoRA for causal language modeling
• Rank: 32
• Alpha: 64
• Dropout: 0.05
• Target modules: q_proj, k_proj, v_proj, o_proj
• Training: 2 epochs, 168 steps, peak learning rate 2e-5
• Run ID: adaption_llama_4_scout_17b_16_final_cc7ddeac

This is an adapter, not a standalone merged model. Load it with PEFT on top of the declared base model and its compatible tokenizer configuration.

Training data

• Hugging Face: https://huggingface.co/datasets/huyxdang/adaption-market-analysis-final
• Kaggle: https://www.kaggle.com/datasets/huydang03/adaption-market-analysis-final

The combined dataset contains 45,758 rows spanning company, fund, macro, fixed-income, and real-estate analysis. Its Adaption data-quality grade improved from B to A.

Evaluation

In Adaption's dataset-specific pairwise preference evaluation, the adapted model was preferred 80 to 20 over the base model. This is a preference result, not an accuracy score. Attempt 2 did not produce a Market Analysis category evaluation, so this result does not establish external-domain transfer.

Limitations

• The adapter inherits the capabilities, restrictions, and risks of its base model.
• The evaluation is dataset-specific and should not be read as investment accuracy.
• Outputs are not financial advice and require human review before real-world use.
• The complete AutoScientist runtime configuration was not exported; this release records the adapter configuration and trainer state included with the run.

Project

• Write-up: https://adaptions-writeup.vercel.app/
• Repository: https://github.com/Kitkitkittt/AdaptMarket
• Authors: Keith Vo and Huy X. Dang

Framework versions

• PEFT 0.15.1