🦊 Toddric-Llama-8B-Merged-v1

Toddric-Llama-8B-Merged-v1 is a merged LoRA adaptation of Meta’s Llama-3.1-8B-Instruct, fine-tuned on custom author-tone data for narrative fluency, creative reasoning, and safe correspondence generation.

The model blends conversational precision with stylistic warmth, avoiding PII leaks and email signature artifacts through cleaned and adversarially-tuned data.
Trained and merged at Foxxe Labs (Ireland) as part of the Toddric-Spidey project.

---

🧠 Training Summary

• Base model: meta-llama/Llama-3.1-8B-Instruct
• Fine-tune method: QLoRA (4-bit, r=64)
• Data: cleaned SFT JSONL set (train_cleaned.jsonl, boosters_cleaned_aligned.jsonl)
• Steps: 300
• Hardware: single RTX 5060 Ti Desktop GPU (16 GB VRAM)
• Frameworks: 🤗 Transformers 4.56 + PEFT 0.10 + bitsandbytes 0.44

---

✅ Key Behaviors

• Retains Toddric writing tone — witty, direct, reflective.
• Produces professional emails and summaries without PII or real-world names.
• Understands authorial context, narrative flow, and technical exposition.
• Refuses to reproduce signatures, URLs, or device footers.

---

🧩 Example Use

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

tok = AutoTokenizer.from_pretrained("toddie314/toddric-llama-8B-merged-v1")
model = AutoModelForCausalLM.from_pretrained("toddie314/toddric-llama-8B-merged-v1",
                                             torch_dtype=torch.bfloat16, device_map="auto")

prompt = "Write a brief, friendly email thanking a conference organizer for their hospitality."
inputs = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=180, temperature=0.7, top_p=0.9)
print(tok.decode(out[0], skip_special_tokens=True))