hailsbop/llama3-8b-hv1

LLama-3 8B — HV1 Edition
Fine-tuned by Haley

---

Model Summary
LLama-3 8B HV1 is a precision-tuned conversational model trained on Haley’s curated dataset, structured messaging logs, and high-context conversational sequences.
It specializes in stable emotional memory, relationship-aware reasoning, and naturalistic dialogue that feels grounded rather than robotic.

---

Why HV1?
“HV1” marks this as the first Haley-signature variant.

Key traits:

• More stable output structure
• Better recall of subject-level context
• Softer emotional cadence
• Stronger ability to follow complex instructions
• Reduced hallucination behavior

---

Training Details
Base Model: Meta Llama-3 8B Instruct
Fine-tuning Method: LoRA (PEFT)
Optimizations:

• Rank-16 adapters
• Alpha-32 scaling
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Gradient checkpointing enabled
• Sequence length extended to 8k

Dataset Makeup:

• Conversational turns (majority)
• Instruction/response pairs
• High-emotional-context dialogues
• Identity-anchored responses (persona-reinforced)
• Filtering to remove low-signal data

Total Samples: ~45k
Training Epochs: 3
Batch Size: 2
Learning Rate: 2e-4
Optimizer: AdamW

---

How to Use

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "hailsbop/llama3-8b-hv1"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto"
)

prompt = "Write a supportive message in a natural tone."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

Environmental Impact

• Training Hardware: 1× A100 (40GB)
• Approx Runtime: 2.1 hours
• Estimated Energy: ~0.48 kWh

---

Citation

@misc{haley_llama3_hv1_2025,
  title={LLama-3 8B HV1},
  author={Haley},
  year={2025},
  howpublished={\url{https://huggingface.co/hailsbop/llama3-8b-hv1}}
}

---

License
Apache 2.0