mouhamadNIbrahim/qwen15-fitness-qlora-adapter

Qwen1.5‑1.8B‑Chat — Fitness & Personal Helper (QLoRA Adapter)

Owner: @mouhamadNIbrahim Base model: Qwen/Qwen1.5-1.8B-Chat Format: PEFT LoRA adapter (trained via QLoRA / 4‑bit NF4) Use case: concise personal helper & evidence‑minded fitness coach

⚠️ Note: This repo contains only the adapter weights. Load it on top of the base model. For serving engines that don’t support PEFT (e.g., vLLM), first merge the adapter (see below).

---

🔰 TL;DR — Quickstart

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
import torch

BASE = "Qwen/Qwen1.5-1.8B-Chat"
ADAPTER = "mouhamadNIbrahim/qwen15-fitness-qlora-adapter"

tok = AutoTokenizer.from_pretrained(BASE)
model = AutoModelForCausalLM.from_pretrained(BASE, torch_dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(model, ADAPTER).to(model.device)

messages = [
    {"role":"system","content":"You are a concise, friendly personal helper and fitness coach. Be evidence-based, practical, and safe."},
    {"role":"user","content":"Give me a 4-day full-body plan and macros for a 69 kg male."}
]
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
outputs = model.generate(**tok([text], return_tensors="pt").to(model.device), max_new_tokens=300)
print(tok.decode(outputs[0], skip_special_tokens=True))

---

🧠 What’s inside

• Finetuned on curated fitness Q&A + helper behaviors (chat format) to give short, actionable responses.
• Keeps Qwen’s ChatML style; training masks only assistant tokens (completion‑only loss).
• LoRA target modules: q_proj, k_proj, v_proj, o_proj, up_proj, gate_proj, down_proj.
• Quantization: bitsandbytes 4‑bit NF4 with double quant (QLoRA recipe).

---

📦 Files

• adapter_model.safetensors — LoRA weights (LFS)
• adapter_config.json — PEFT/LoRA config
• (optional) tokenizer files — not required for adapters; base model tokenizer is recommended
• chat_template.jinja — ChatML template reference

---

🛠️ Merge to a single model (optional)

Some runtimes (e.g., vLLM, TGI) need merged weights.

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
import torch

BASE = "Qwen/Qwen1.5-1.8B-Chat"
ADPT = "mouhamadNIbrahim/qwen15-fitness-qlora-adapter"

tok = AutoTokenizer.from_pretrained(BASE)
base = AutoModelForCausalLM.from_pretrained(BASE, torch_dtype=torch.bfloat16, device_map="auto")
merged = PeftModel.from_pretrained(base, ADPT).merge_and_unload()
merged.save_pretrained("./qwen15-fitness-merged", safe_serialization=True)
tok.save_pretrained("./qwen15-fitness-merged")

---

📚 Training details

• Base: Qwen/Qwen1.5-1.8B-Chat
• Method: QLoRA (PEFT/LoRA on 4‑bit base)
• Tokenizer: base model tokenizer + apply_chat_template
• Loss: completion‑only (mask non‑assistant tokens)
• LoRA: r=16, alpha=32, dropout=0.05, bias=none
• Targets: q_proj,k_proj,v_proj,o_proj,up_proj,gate_proj,down_proj
• Optimizer: paged_adamw_8bit
• Scheduler: cosine; warmup 3%
• Precision: bf16 if available (fallback fp16)
• Seq length: up to 2048 tokens
• Batching: per_device_train_batch_size=2, gradient_accumulation_steps=8
• Other: gradient checkpointing enabled
• Hardware: single GPU (tested on RTX 4090 / A4000 / 3090 class)

Reproducibility: see training_args.bin and trainer_state.json in your training runs. Dataset not released; marked as private.

---

🧪 Evaluation (placeholder)

No formal benchmarks included. Manually validated on:

• Workout planning prompts (4‑day/5‑day splits, full‑body, push/pull/legs)
• Macro guidance and calorie targets (Mifflin‑St Jeor, TDEE assumptions)
• Lifestyle coaching quick answers

If you run quantitative evals (e.g., custom fitness QA set), please share results via PR/issues.

---

⚖️ Intended use & limitations

• Intended use: consumer fitness guidance, scheduling/helpful assistant tasks, basic nutrition coaching.
• Not for: medical diagnosis, treatment, or individualized clinical nutrition. The model can be confident yet wrong.
• Use responsibly: always verify critical advice; consult a professional for injuries/health conditions.

---

🔒 Safety

The finetune includes examples that: request context (age, weight, goals), caution on injuries, recommend professional help when necessary. Still, errors or bias can occur.

---

📝 Example prompts

• “Design a 4‑day full‑body plan for a beginner with access to dumbbells only.”
• “I’m 69 kg, 173 cm, ~8k steps/day. Estimate maintenance calories and macros for lean bulk.”
• “Give me 3 high‑protein Lebanese meal ideas under 700 kcal each.”

---

🔁 Training recipe (script sketch)

Uses transformers, peft, trl, bitsandbytes.

# Key ingredients used in training
from transformers import BitsAndBytesConfig
from peft import LoraConfig
from trl import SFTTrainer, DataCollatorForCompletionOnlyLM

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
)

lora = LoraConfig(
    r=16, lora_alpha=32, lora_dropout=0.05, bias="none",
    task_type="CAUSAL_LM",
    target_modules=["q_proj","k_proj","v_proj","o_proj","up_proj","gate_proj","down_proj"],
)

# Data formatted with tokenizer.apply_chat_template([...])
# and collated with DataCollatorForCompletionOnlyLM(
#   response_template="<|im_start|>assistant\n", tokenizer=tokenizer
# )

---

📥 How to cite

@software{qwen15_fitness_qlora_adapter,
  author = {Ibrahim, Mouhamad},
  title  = {Qwen1.5-1.8B-Chat — Fitness & Personal Helper (QLoRA Adapter)},
  year   = {2025},
  url    = {https://huggingface.co/mouhamadNIbrahim/qwen15-fitness-qlora-adapter}
}

---

🙏 Acknowledgements

• QLoRA: Dettmers et al. (2023)
• PEFT: Mangrulkar et al. (2022)
• Qwen team for the base model and ChatML template

---

📄 License

• Adapter released under Apache‑2.0. The base model’s original license also applies when combining weights.