Training Procedure

Hardware:

4x NVIDIA A100-SXM4-80GB (DDP via torchrun --nproc_per_node=4)

LoRA config:

rank r=32
alpha 64
dropout 0.05
target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj

Optimization:

epochs: 3
learning rate: 2e-4
per-device train batch size: 4
gradient accumulation: 4
max sequence length: 1024
mixed precision: bf16
optimizer: paged AdamW 8-bit

Metrics Snapshot

From trainer_state.json and train_metrics.json:

train points: 45
eval points: 9
train loss: 4.12497 -> 0.10813
eval loss: 2.06111 -> 0.11857
best checkpoint: checkpoint-45
train runtime: 88.83s
train run average loss: 0.69445

Quick Start

Transformers + PEFT

python

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

BASE = "Qwen/Qwen3-0.6B"
ADAPTER = "mr-dee/qwen3-ascii-tui-lora"

tokenizer = AutoTokenizer.from_pretrained(BASE, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    BASE,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(model, ADAPTER)

messages = [
    {"role": "system", "content": "You create terminal-friendly educational ASCII/TUI diagrams."},
    {"role": "user", "content": "Explain the double-slit experiment with a compact ASCII diagram."},
]
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=600, temperature=0.2, top_p=0.9)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

vLLM (OpenAI-compatible endpoint)

bash

vllm serve Qwen/Qwen3-0.6B \
  --enable-lora \
  --lora-modules ascii=mr-dee/qwen3-ascii-tui-lora \
  --tensor-parallel-size 2 \
  --max-model-len 2048

Query example:

bash

curl http://127.0.0.1:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "ascii",
    "messages": [
      {"role":"user","content":"Teach gravity using an ASCII flow diagram and 3 takeaways."}
    ],
    "max_tokens": 500,
    "temperature": 0.2
  }'

Important:

keep prompt_tokens + max_tokens <= max_model_len
if exceeded, vLLM returns HTTP 400 validation errors

Risks and Limitations

May generate plausible but incorrect scientific claims.
ASCII layout quality can degrade for very long outputs.
Not guaranteed to follow strict width limits unless prompted.
Should be reviewed by a human before educational publication.

Responsible Use

Verify scientific correctness.
Validate generated text/code before classroom use.
Avoid over-reliance in high-stakes domains.

Framework Versions

PEFT: 0.18.1
Transformers: 4.57.6
TRL: 0.27.2
PyTorch: 2.9.1

Citation

bibtex

@misc{qwen3_ascii_tui_lora_2026,
  title  = {qwen3-ascii-tui-lora},
  author = {mr-dee},
  year   = {2026},
  url    = {https://huggingface.co/mr-dee/qwen3-ascii-tui-lora}
}