80 lines
2.0 KiB
Markdown
80 lines
2.0 KiB
Markdown
---
|
|
language:
|
|
- en
|
|
license: other
|
|
library_name: transformers
|
|
pipeline_tag: text-generation
|
|
base_model: Qwen/Qwen3-4B-Thinking-2507
|
|
base_model_relation: finetune
|
|
tags:
|
|
- qwen
|
|
- qwen3
|
|
- reasoning
|
|
- chain-of-thought
|
|
- text-generation
|
|
- bf16
|
|
- merged
|
|
datasets:
|
|
- HuggingFaceH4/Bespoke-Stratos-17k
|
|
- nvidia/OpenMathReasoning
|
|
- nvidia/OpenCodeReasoning
|
|
- nvidia/OpenScienceReasoning-2
|
|
- lordx64/reasoning-distill-opus-4-7-max-sft
|
|
---
|
|
|
|
# Qwen3-4B-MegaR3ASONER-v1
|
|
|
|
A full merged reasoning model created by merging the MegaR3ASONER LoRA into
|
|
[`Qwen/Qwen3-4B-Thinking-2507`](https://huggingface.co/Qwen/Qwen3-4B-Thinking-2507).
|
|
|
|
## Model type
|
|
|
|
This repository contains the complete merged Transformers model, not only the
|
|
PEFT adapter. It can be loaded directly without `PeftModel`.
|
|
|
|
- Base model: `Qwen/Qwen3-4B-Thinking-2507`
|
|
- Merge dtype: BF16
|
|
- Adapter source: `RexTRO111/Qwen3-4B-MegaR3ASONER-LoRA-v1`
|
|
- Training hardware: NVIDIA A10G on Modal
|
|
- Merge hardware: NVIDIA A10G on Modal
|
|
|
|
## 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 a full GSM8K score and not a
|
|
controlled base-versus-fine-tune comparison.
|
|
|
|
## Loading
|
|
|
|
```python
|
|
import torch
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
|
|
model_id = "RexTRO111/Qwen3-4B-MegaR3ASONER-v1"
|
|
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
|
model = AutoModelForCausalLM.from_pretrained(
|
|
model_id,
|
|
torch_dtype=torch.bfloat16,
|
|
device_map="auto",
|
|
)
|
|
model.eval()
|
|
```
|
|
|
|
## Limitations
|
|
|
|
- It may produce an incorrect intermediate thought before correcting itself.
|
|
- It can overthink simple prompts.
|
|
- Long reasoning traces increase latency and cost.
|
|
- Benchmark contamination has not been exhaustively ruled out.
|
|
- Verify answers before high-stakes use.
|
|
|
|
## Licensing note
|
|
|
|
The Qwen base model and every training dataset retain their own licenses and
|
|
upstream terms. Review all applicable terms before redistribution or
|
|
commercial use.
|