GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Vijayendra
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Vijayendra
Model Tree
openai/gpt-oss-20b
Input Modalities
Output Modalities
Supported Functionality
Training uses the official training splits from:
GPQA is intentionally not used for SFT so that GPQA/GPQA-Diamond can remain clean external evaluation benchmarks.
Total source training examples before max-length filtering: 17,755.
GPT-OSS's native Harmony chat template is used. System/user tokens are masked from the LM objective and training supervises the assistant response.
When source datasets provide an evidence passage/fact, it is included as
Supporting evidence: in the final assistant response. No synthetic
chain-of-thought was fabricated for this training set.
openai/gpt-oss-20bpython
from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel base_id = "openai/gpt-oss-20b"adapter_id = "Vijayendra/gpt-oss-20b-SciReason-LoRA" tokenizer = AutoTokenizer.from_pretrained(adapter_id)base = AutoModelForCausalLM.from_pretrained( base_id, device_map="auto", dtype="auto", attn_implementation="eager",)model = PeftModel.from_pretrained(base, adapter_id)The small sanity evaluation uploaded with the adapter is not intended to replace the final benchmark suite. ARC-Challenge, OpenBookQA, SciQ, GPQA/GPQA-Diamond, and out-of-domain controls should be reported separately with fixed evaluation protocols.
The GPT-OSS base model is Apache-2.0. Dataset use remains subject to each upstream dataset's license and terms.