Model Card: Gemma3-1B Turkish CPT (150K–200K Subset, 1 Epoch – Stage 4)
Overview
This model is the Stage 4 Turkish Continued Pretraining (CPT) variant of Gemma-3-1B.
Unlike Stage 1, which was initialized from google/gemma-3-1b-pt,
this model was initialized from:
canbingol/gemma3_1B_base-tr-cpt-1epoch_stage3
Stage 4 continues domain adaptation by exposing the model to new data rather than repeating the same subset.
The model was trained for 1 epoch on samples 150,000 to 200,000 of the Turkish web corpus.
Importantly, this model is a direct continuation of Stage 3.
Therefore, cumulatively it has been trained on samples 0–200,000 of the corpus (Stage 1: 0–50K, Stage 2: 50K–100K, Stage 3: 100K–150K, Stage 4: 150K–200K).
This stage corresponds to the end of the 1-epoch pass over the full 200K-sample dataset (i.e., completion of the first full epoch via sequential shards).
Training Lineage
- Stage 0:
google/gemma-3-1b-pt
- Stage 1: Samples 0–50,000 (1 epoch)
- Stage 2: Samples 50,000–100,000 (1 epoch)
- Stage 3: Samples 100,000–150,000 (1 epoch)
- Stage 4 (this release): Samples 150,000–200,000 (1 epoch, end of epoch-1)
Cumulative data exposure: 0–200,000 samples
This represents sequential CPT across disjoint data shards.
Training Setup
- Dataset:
canbingol/vngrs-web-corpus-200k
- Subset Used: Samples 150,000–200,000
- Initialization: Stage 3 checkpoint
- Training Objective: Continued Pretraining
- Epochs: 1
- Data Regime: Plain text
- Token Count: ~21.6M tokens
- Cumulative Token Exposure (Stage 1 + Stage 2 + Stage 3 + Stage 4): ~86.1M tokens (approximate)
Notes on cumulative exposure:
- Although Stage 4 trains only on the 150K–200K shard, it inherits all adaptations learned from previous stages.
- After this stage, the model has effectively completed exposure to the entire 0–200K dataset range through sequential continuation.
Usage Example
1import torch
2from transformers import AutoModelForCausalLM, AutoTokenizer
3
4model_id = "canbingol/gemma3_1B_base-tr-cpt-1epoch_stage4"
5
6device = "cuda" if torch.cuda.is_available() else "cpu"
7
8model = AutoModelForCausalLM.from_pretrained(model_id)
9tokenizer = AutoTokenizer.from_pretrained(model_id)
10
11model = model.to(device)
12
13prompt = "bundan böyle"
14inputs = tokenizer(prompt, return_tensors="pt").to(device)
15
16outputs = model.generate(
17 **inputs,
18 max_new_tokens=50,
19 do_sample=True,
20 temperature=0.8,
21 top_p=0.9
22)
23
24generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
25print(generated_text)
