Repo: michael-sigamani/llama2-7b-tat-lora-fp16
Base: NousResearch/Llama-2-7b-hf
Adapter: next-tat/tat-llm-7b-lora
Merged: ✅ Yes
Fine-tuned: ❌ Not yet (this is the pre-CoT stage)
Format: Float16 (fp16)

📖 Overview

This model merges a FinQA-tuned adapter (TAT) into LLaMA 2 7B, producing a standalone checkpoint ready for further fine-tuning or inference on financial reasoning tasks.

📈 Finetuned LoRA (TAT) captures scalar reasoning from FinQA
🔁 Merged via peft.merge_and_unload() into the full model
🧵 Next step: fine-tune on train_turn.jsonl with chain-of-thought (CoT) supervision

🔁 Intended Usage

Use this model as the starting point for:

🧠 Fine-tuning on CoT financial datasets (e.g. ConvFinQA turn-based reasoning)
🧪 Evaluation on scalar, program, and reasoning benchmarks
🦙 Export to GGUF for Ollama / llama.cpp

🚧 Not a Final Model

This checkpoint has not been CoT fine-tuned yet. It is the output of:

bash

Base:   NousResearch/Llama-2-7b-hf
LoRA:   next-tat/tat-llm-7b-lora (FinQA-style)
Merged: Yes (fp16, no adapter required)

Next step: Train on chain-of-thought examples (train_turn.jsonl) using Unsloth or PEFT + TRL.

🧠 Merge Script

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained("NousResearch/Llama-2-7b-hf")
adapter = PeftModel.from_pretrained(base_model, "next-tat/tat-llm-7b-lora")
merged = adapter.merge_and_unload()

merged.save_pretrained("llama2-7b-tat-lora-fp16")
AutoTokenizer.from_pretrained("NousResearch/Llama-2-7b-hf").save_pretrained("llama2-7b-tat-lora-fp16")

🧑‍💻 Maintainer

Michael Sigamani
github.com/sigamani

📜 License

Base: Meta LLaMA 2 license (via NousResearch)
Adapter: Apache 2.0
Merged model: Inherits original LLaMA 2 license – requires HF auth

michael-sigamani/llama2-7b-tat-convfinqa-fp16

🦙 LLaMA 2 7B + TAT FinQA Adapter (Merged, Pre-CoT)