38 lines
862 B
Markdown
38 lines
862 B
Markdown
---
|
|
license: apache-2.0
|
|
library_name: transformers
|
|
pipeline_tag: text-generation
|
|
tags:
|
|
- qwen3
|
|
- qwen
|
|
- fine-tuned
|
|
base_model: Qwen/Qwen3-4B-Base
|
|
---
|
|
|
|
# Qwen3-4B No-Think Path3 Checkpoint-2500
|
|
|
|
Deliverable merged BF16 checkpoint for the Qwen3-4B no-think Path3 training run at checkpoint 2500.
|
|
|
|
## Files
|
|
|
|
- `model.safetensors` — merged BF16 weights
|
|
- `config.json` / `generation_config.json`
|
|
- `tokenizer.json` / `tokenizer_config.json`
|
|
- `chat_template.jinja`
|
|
|
|
## Load
|
|
|
|
```python
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
|
|
repo = "modrill/Qwen3-4B-nothink-path3-ckpt2500"
|
|
tok = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
|
|
model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="bfloat16", device_map="auto", trust_remote_code=True)
|
|
```
|
|
|
|
## Notes
|
|
|
|
- Precision: BF16
|
|
- Checkpoint: 2500
|
|
- Training path: no-think Path3
|