GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
aishwaryawambule
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
aishwaryawambule
Model Tree
openai-community/gpt2
Input Modalities
Output Modalities
Supported Functionality
GPT-2 small (124M parameters, frozen) fine-tuned for movie-review sentiment (SST-2). The circuit generates the weight changes, so training adjusts only 1,692 circuit angles (plus a 1,536-number classification head). LoRA rank 4 would train 147,456 numbers for the same layers.
| Accuracy (SST-2 validation, 872 reviews) | 86.7% (756 / 872) |
| Trainable numbers | 3,228 (1,692 circuit angles + 1,536 head) |
| Adapter file size | 15 KB |
| Circuit per layer | 14 qubits, 10 rounds of RY rotations + a CNOT ring |
bash
pip install quantum-adapterhf download aishwaryawambule/gpt2-sst2-quantum-adapter --local-dir sst2-adapterquantum-adapter predict --adapter sst2-adapter "A beautiful, moving film."# positive 100% A beautiful, moving film.Or from Python:
python
from transformers import AutoModelForSequenceClassification, AutoTokenizerfrom quantum_adapter import QuantumModel tok = AutoTokenizer.from_pretrained("gpt2")tok.pad_token = tok.eos_tokenbase = AutoModelForSequenceClassification.from_pretrained("gpt2", num_labels=2)base.config.pad_token_id = tok.pad_token_idmodel = QuantumModel.from_pretrained(base, "aishwaryawambule/gpt2-sst2-quantum-adapter") logits = model(**tok(["A beautiful, moving film."], return_tensors="pt")).logitsprint(["negative", "positive"][logits.argmax(-1).item()])With the library's own command, on 4,000 SST-2 training sentences (10% held out for picking the best epoch):
bash
quantum-adapter train --csv sst2_train.csv --out sst2-adapter --model gpt2 --lr 3e-3 --seed 0The training output (on an Apple M5 Pro, about 8 minutes):
text
4000 rows, 2 labels ['negative', 'positive']: training on 3601, validating on 399.trainable params: 3,228 || all params: 124,443,036 || trainable%: 0.0026epoch 1/6: train loss 0.6641 validation accuracy 75.4%epoch 2/6: train loss 0.4193 validation accuracy 82.2%epoch 3/6: train loss 0.3827 validation accuracy 84.7%epoch 4/6: train loss 0.3530 validation accuracy 83.2%epoch 5/6: train loss 0.3509 validation accuracy 83.2%epoch 6/6: train loss 0.3313 validation accuracy 81.2%kept epoch 3 (best validation accuracy 84.7%)Saved the adapter to sst2-adapter/. Try it: quantum-adapter predict --adapter sst2-adapter "some text"6 epochs, batch size 32, circuit learning rate 3e-3, head learning rate 1e-3, no weight decay. Validation accuracy on the 399 held-out rows peaked at epoch 3, so that epoch is the saved adapter. It was then scored on the 872 SST-2 validation reviews it never saw: 86.7% (756 / 872).
Two predictions from the saved adapter:
text
$ quantum-adapter predict --adapter sst2-adapter "A beautiful, moving film." "Two hours of my life I will never get back."positive 100% A beautiful, moving film.negative 84% Two hours of my life I will never get back.See the box at the top, and the README for every option.
pip install quantum-adapterhf download aishwaryawambule/gpt2-sst2-quantum-adapter --local-dir sst2-adapterquantum-adapter predict --adapter sst2-adapter "A beautiful, moving film."# positive 100% A beautiful, moving film.from transformers import AutoModelForSequenceClassification, AutoTokenizerfrom quantum_adapter import QuantumModel tok = AutoTokenizer.from_pretrained("gpt2")tok.pad_token = tok.eos_tokenbase = AutoModelForSequenceClassification.from_pretrained("gpt2", num_labels=2)base.config.pad_token_id = tok.pad_token_idmodel = QuantumModel.from_pretrained(base, "aishwaryawambule/gpt2-sst2-quantum-adapter") logits = model(**tok(["A beautiful, moving film."], return_tensors="pt")).logitsprint(["negative", "positive"][logits.argmax(-1).item()])quantum-adapter train --csv sst2_train.csv --out sst2-adapter --model gpt2 --lr 3e-3 --seed 04000 rows, 2 labels ['negative', 'positive']: training on 3601, validating on 399.trainable params: 3,228 || all params: 124,443,036 || trainable%: 0.0026epoch 1/6: train loss 0.6641 validation accuracy 75.4%epoch 2/6: train loss 0.4193 validation accuracy 82.2%epoch 3/6: train loss 0.3827 validation accuracy 84.7%epoch 4/6: train loss 0.3530 validation accuracy 83.2%epoch 5/6: train loss 0.3509 validation accuracy 83.2%epoch 6/6: train loss 0.3313 validation accuracy 81.2%kept epoch 3 (best validation accuracy 84.7%)Saved the adapter to sst2-adapter/. Try it: quantum-adapter predict --adapter sst2-adapter "some text"$ quantum-adapter predict --adapter sst2-adapter "A beautiful, moving film." "Two hours of my life I will never get back."positive 100% A beautiful, moving film.negative 84% Two hours of my life I will never get back.