Meghan

Meghan is a chat-tuned Qwen3-4B, built as a friendly "classmate AI" study buddy trained on the public course catalog site of a US public high school plus authored study walkthroughs. She knows every course in the catalog (titles, numbers, credits, topics, prerequisites, department structure), common class policies, and how to walk math and science problems through step by step in a casual teen-adjacent tone. The training pipeline deliberately keeps the source school anonymous.

Lineage and training

• Base model: Qwen3-4B (via the unsloth 4-bit checkpoint)
• Method: QLoRA (r=16, alpha=16), 3 epochs, effective batch 16, lr 2e-4 cosine, bf16
• Data: 1,403 chat examples (1262 train / 141 holdout):
  • ~53% catalog knowledge Q&A derived from the school's public course pages and PDF Program of Studies
  • ~25% worked solver walkthroughs with a verified ```calc tool-block convention
  • ~13% persona / casual conversation
  • ~9% page-context grounding tasks
• All arithmetic in the solver dataset is programmatically audited (each expr = result line evaluates true)
• No student data, no private district data: corpus is 100% public-website content + authored examples

Files

Path	What
meghan-Q4_K_M.gguf	4-bit quantized GGUF (~2.4 GB) for llama.cpp/Ollama/Jan
model-0000X-of-00002.safetensors	merged bf16 weights (~7.7 GB) for transformers/vLLM
adapter/	the raw QLoRA adapter (132 MB) if you want to Mount on your own base

Usage (transformers)

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("sauceyy909/Meghan")
model = AutoModelForCausalLM.from_pretrained("sauceyy909/Meghan", torch_dtype="bfloat16", device_map="auto")
msgs = [{"role": "user", "content": "What topics does Algebra 2 cover?"}]
prompt = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
out = model.generate(**tok(prompt, return_tensors="pt").to(model.device), max_new_tokens=300)
print(tok.decode(out[0], skip_special_tokens=True))

Usage (llama.cpp / Ollama)

ollama run sauceyy909/Meghan:Q4_K_M
# or llama-quantize'd GGUF via llama-cli -m meghan-Q4_K_M.gguf

Caveats

• Small model, small data: treat as a study helper, not an authority. Verify anything load-bearing.
• Course facts describe one (anonymous) US public high school curriculum as published publicly.
• License: follows Qwen3 base terms (Qwen Research / Agreement) — redistribute under compatible terms.