AlihanSDev/kindlehare-qwen7b-lora

KindleHare — Qwen2.5-7B LoRA

QLoRA fine-tune of Qwen/Qwen2.5-7B-Instruct on agentic reasoning traces.

Model Details

• Base model: Qwen/Qwen2.5-7B-Instruct
• Fine-tuning method: QLoRA (4-bit NF4 + LoRA r=16)
• Developed by: AlihanSDev
• License: apache-2.0

Training Data

Dataset: lambda/hermes-agent-reasoning-traces (kimi split)

• 7646 samples of agentic coding and reasoning traces
• Categories: Terminal Tasks, coding, reasoning
• Filtered: tool/system messages removed, only user↔assistant turns kept

Training Details

Parameter	Value
Hardware	1x NVIDIA T4 (Kaggle)
Epochs	1
Learning rate	1e-4
Batch size	1 (grad acc 8)
LoRA rank	16
LoRA alpha	16
Target modules	q/k/v/o/gate/up/down proj
Precision	fp16
Optimizer	paged_adamw_8bit
Max sequence length	1024
Training time	~8 hours

How to Use

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

base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-7B-Instruct",
    torch_dtype=torch.float16,
    device_map="auto"
)
model = PeftModel.from_pretrained(base_model, "AlihanSDev/kindlehare-qwen7b-lora")
tokenizer = AutoTokenizer.from_pretrained("AlihanSDev/kindlehare-qwen7b-lora")

messages = [{"role": "user", "content": "Write a Python function to reverse a string."}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
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))

Limitations

• Trained for 1 epoch only
• Eval during training was disabled due to T4 memory constraints
• Loss was noisy due to diverse complexity of agentic traces in dataset
• Not evaluated on standard benchmarks (MMLU, HumanEval)

Framework versions

• PEFT 0.12.0