Important
This repository contains LoRA adapter weights only. Load it on the exact
base checkpoint:
Qwen/Qwen3-4B-Thinking-2507
Do not attach it to Qwen/Qwen3-4B-Instruct-2507 or another base model.
Training overview
- Method: BF16 16-bit LoRA
- Hardware: NVIDIA A10G on Modal
- Focus: mathematical, scientific, coding, and general reasoning
- Training sources:
HuggingFaceH4/Bespoke-Stratos-17knvidia/OpenMathReasoningnvidia/OpenCodeReasoningnvidia/OpenScienceReasoning-2lordx64/reasoning-distill-opus-4-7-max-sft
Preliminary evaluation
On the first 100 examples selected by EleutherAI's gsm8k_cot task:
- Flexible extraction exact match: 88%
- Strict match: 83%
This was a limited 100-question run, not the complete GSM8K test set and not a
controlled base-versus-adapter comparison.
Loading
import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base_id = "Qwen/Qwen3-4B-Thinking-2507"
adapter_id = "RexTRO111/Qwen3-4B-MegaR3ASONER-LoRA-v1"
tokenizer = AutoTokenizer.from_pretrained(base_id)
base = AutoModelForCausalLM.from_pretrained(
base_id,
torch_dtype=torch.bfloat16,
device_map="auto",
)
model = PeftModel.from_pretrained(
base,
adapter_id,
is_trainable=False,
)
model.eval()
Limitations
- The model may produce incorrect intermediate reasoning even when its final
answer is correct.
- It can overthink simple prompts.
- Long reasoning traces increase latency.
- Benchmark contamination has not been exhaustively ruled out.
- Verify outputs before high-stakes use.
Licensing note
The base model and each training dataset retain their own licenses and upstream
terms. Review all of them before redistribution or commercial use.