kunjcr2/stablelm-1.6b-finetuned-aligned

StableLM-1.6B Fine-Tuned & Aligned (SFT + GRPO)

This repository contains a LoRA-based aligned variant of StableLM-2-1.6B, trained through a complete post-pretraining pipeline consisting of Supervised Fine-Tuning (SFT) followed by GRPO-based alignment.

This is a learning / portfolio model demonstrating end-to-end post-pretraining, not a production-ready system.

---

Model Overview

• Base model: stabilityai/stablelm-2-1_6b
• Model type: PEFT / LoRA adapter
• Parameters (base): ~1.64B
• Trainable parameters (LoRA): ~30–60M (depending on adapter stack)
• License: Apache 2.0

⚠️ Important: This repository contains LoRA adapter weights only (adapter_model.safetensors). You must load it on top of the base StableLM model.

---

Training Pipeline

Base StableLM-2-1.6B
        │
        ▼
Supervised Fine-Tuning (LoRA, UltraChat)
        │
        ▼
GRPO Alignment (Reward-guided, PKU-SafeRLHF)
        │
        ▼
Aligned LoRA Adapter (this repo)

---

Stage 1: Supervised Fine-Tuning (SFT)

• Objective: Instruction-following capability

• Method: LoRA-based fine-tuning

• Dataset: HuggingFaceH4/ultrachat_200k

• LoRA configuration:

  • Applied to attention and MLP projections
  • Parameter-efficient (base model frozen)

---

Stage 2: GRPO Alignment

• Objective: Safety and preference alignment
• Method: Group Relative Policy Optimization (GRPO)
• Reward model: OpenAssistant/reward-model-deberta-v3-large
• Dataset: PKU-Alignment/PKU-SafeRLHF

Training Metrics

The following metrics were tracked during GRPO alignment:

• Train Loss: Stable around zero, indicating controlled policy updates
• Reward: Increases steadily and saturates, showing successful alignment
• Entropy: Decreases significantly, reflecting reduced policy randomness and convergence

Github: https://github.com/kunjcr2/end-to-end-post-pretraining

📊 Weights & Biases run:

https://wandb.ai/kunjcr2-dreamable/huggingface/runs/qds5kqi2?nw=nwuserkunjcr2

---

How to Load the Model

This model cannot be loaded standalone.

Correct loading

from transformers import AutoModelForCausalLM
from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained(
    "stabilityai/stablelm-2-1_6b",
    trust_remote_code=True
)

model = PeftModel.from_pretrained(
    base_model,
    "kunjcr2/stablelm-1.6b-finetuned-aligned"
)

The tokenizer included in this repository should be used to ensure consistency with training.

---

Intended Use

• Demonstrating an end-to-end LLM post-pretraining pipeline
• Research, experimentation, and educational purposes
• Portfolio / reference implementation for SFT + alignment workflows

---

Limitations

• Not production-grade
• Limited training steps and data scale
• Alignment quality depends on reward model biases
• No extensive safety evaluation beyond training metrics

---

Citation

If you reference this work, please cite the base model and datasets:

• StableLM-2-1.6B — Stability AI
• UltraChat 200K — Hugging Face H4
• PKU-SafeRLHF — PKU Alignment
• OpenAssistant Reward Model

---

License

Apache 2.0. See the LICENSE file for details.