Base model: danielhanchen/gpt-oss-20b
Training method: semi-online TTRL with vLLM rollout collection and reward-filtered QLoRA updates
CSV schema used: id, category, problem, answer, source
Topics processed: Algebra, Number Theory
Best recorded eval accuracy: not recorded
Mean recorded rollout reward: 0.7031
Model card generated at: 2026-04-05 13:02:06 UTC
Template version: 2026-04-04

Artifact Details

Artifact type: LoRA/QLoRA adapter
Latest adapter path: ttrl_artifacts/adapters/algebra_batch_000

Usage

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model_name = 'danielhanchen/gpt-oss-20b'
adapter_repo_id = 'AnthonyNwafor/gpt-oss-20b-ttrl-adapter'

tokenizer = AutoTokenizer.from_pretrained(base_model_name, trust_remote_code=True)
base_model = AutoModelForCausalLM.from_pretrained(base_model_name, trust_remote_code=True)
model = PeftModel.from_pretrained(base_model, adapter_repo_id)

Notes

Rollouts were collected with a local vLLM OpenAI-compatible server.
Training data provenance should be tracked separately alongside the rollout JSONL files and summary CSV.
If you publish benchmark numbers, keep held-out evaluation separate from any adaptation pool.