File size: 1,784 Bytes
ae6e8cb
 
 
384a764
 
 
61f8c85
384a764
 
 
 
61f8c85
384a764
 
 
61f8c85
384a764
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
---
license: cc-by-nc-4.0
---

# Qwen3-1.7B on Nemotron-CC (low-quality tier): ablation checkpoints

Four base language models with the Qwen3-1.7B architecture (1.72B parameters). Each was pretrained from scratch for one epoch on 34.4B tokens of the low-quality (LQ) tier of Nemotron-CC, Common Crawl snapshots 2013–2015.

| Folder | Training data | Seed |
|---|---|---|
| `v2-lq-seed1234/` | LQ after V2 repair (prose-only repair + linguistic repair) | 1234 |
| `raw-lq-seed1234/` | LQ, unrepaired | 1234 |
| `raw-lq-seed1337/` | LQ, unrepaired | 1337 |
| `raw-lq-seed42/` | LQ, unrepaired | 42 |

- **Seeds:** the seed sets the weight initialization only. Every run used the same data-shuffling seed (1234), so the three raw runs saw identical batches in identical order. `v2-lq-seed1234/` and `raw-lq-seed1234/` share the same configuration and seed and differ only in training data.
- **Training:** 16,406 steps × 512 sequences × 4,096 tokens, BF16, AdamW (β = 0.9/0.95, weight decay 0.1), learning rate 3e-4 → 3e-5 (warmup–stable–decay).
- **Base models only:** no instruction tuning.

## Format and compatibility

Each folder is a standard Hugging Face checkpoint: BF16 `model.safetensors`, tokenizer, and generation config. The checkpoints were exported with transformers 5.3. `config.json` and `tokenizer_config.json` also carry the older field names, so transformers 4.x (4.51 or newer) loads the same RoPE base (1,000,000) and the same tokenizer.

## Usage

```bash
hf download anonymous-md/test --include "v2-lq-seed1234/*" --local-dir ./qwen3-lq
```

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

path = "./qwen3-lq/v2-lq-seed1234"
tokenizer = AutoTokenizer.from_pretrained(path)
model = AutoModelForCausalLM.from_pretrained(path)
```