GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
sagnikdas-ml
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
sagnikdas-ml
Model Tree
google/gemma-4-12B-it
Input Modalities
Output Modalities
Supported Functionality
Use this adapter with its exact base model for educational question answering, preferably with retrieval and citation validation. It is not a standalone model and does not contain the base weights.
Do not treat its answers as authoritative scholarship, grading decisions, or professional advice. The adapter is not designed to reproduce readings or to answer unsupported questions from memory.
python
import torchfrom peft import PeftModelfrom transformers import AutoModelForMultimodalLM, AutoProcessor, BitsAndBytesConfig base_model = "google/gemma-4-12B-it"adapter = "sagnikdas-ml/epistemology-gemma4-12b-v0.1" processor = AutoProcessor.from_pretrained(base_model)model = AutoModelForMultimodalLM.from_pretrained( base_model, device_map="auto", dtype=torch.bfloat16, quantization_config=BitsAndBytesConfig( load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_compute_dtype=torch.bfloat16, bnb_4bit_use_double_quant=True, ),)model = PeftModel.from_pretrained(model, adapter)model.eval()Access to the gated base model may require accepting Google's Gemma terms on Hugging Face and authenticating locally.
The adapter was trained with supervised fine-tuning on 16 approved examples derived from course concepts. Examples carry provenance metadata and cover definitions, explanations, comparisons, argument reconstruction/evaluation, objections, applications, misconception correction, and unsupported-question behavior. The original course documents are not distributed with this model.
| Setting | Value |
|---|---|
| Method | QLoRA / supervised fine-tuning |
| LoRA rank / alpha / dropout | 16 / 32 / 0.05 |
| Target modules | Attention and MLP projection layers |
| Quantization | NF4, double quantization, BF16 compute |
| Epochs | 3 |
| Learning rate | 2e-4, cosine schedule |
| Maximum sequence length | 2048 |
| Effective batch size | 32 |
| Seed | 42 |
| Hardware | 1× NVIDIA A40 |
Evaluation used a source-separated validation split of 2 examples.
| Metric | Result |
|---|---|
| Validation loss | 4.1711 |
| Mean token accuracy | 0.3583 |
| Validation tokens | 6,840 |
These figures are diagnostic rather than a robust estimate of generalization: the validation set is extremely small. An end-to-end smoke test with the course RAG system produced a passage-grounded answer with valid citations, but no broad human evaluation has yet been completed.
For consequential use, verify answers against primary texts and qualified instructors. Keep retrieval restricted to materials you are authorized to use.
The training configuration, run manifest, validation metrics, inference code,
and tests are available in the
Epistemology-Coursebot
repository. Key library versions were PyTorch 2.12.1, Transformers 5.12.1,
PEFT 0.19.1, TRL 1.13.0, and BitsAndBytes 0.49.2.
import torchfrom peft import PeftModelfrom transformers import AutoModelForMultimodalLM, AutoProcessor, BitsAndBytesConfig base_model = "google/gemma-4-12B-it"adapter = "sagnikdas-ml/epistemology-gemma4-12b-v0.1" processor = AutoProcessor.from_pretrained(base_model)model = AutoModelForMultimodalLM.from_pretrained( base_model, device_map="auto", dtype=torch.bfloat16, quantization_config=BitsAndBytesConfig( load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_compute_dtype=torch.bfloat16, bnb_4bit_use_double_quant=True, ),)model = PeftModel.from_pretrained(model, adapter)model.eval()