---
license: apache-2.0
base_model: unsloth/Qwen3.5-4B-Base
tags: [tajik, qwen3.5, continual-pretraining, cpt, merged, stacked-lora]
language: [tg]
---

# Qwen3.5-4B Tajik — CPT epoch-2 (stacked), merged

Continual-pretraining (CPT) adaptation of **Qwen3.5-4B-Base** to **Tajik** (тоҷикӣ), **2 stacked epochs**,
with the LoRA adapters merged into the base weights (BF16). This is the **stacked-LoRA** chain:
epoch-2 is a *fresh* rank-64 LoRA trained on top of the merged **base+e1** checkpoint
(`Tohirju/qwen35-4b-tajik-cpt-e1-merged`), then merged in — i.e. `base + e1 + e2`.

- **What it is:** Qwen3.5-4B-Base after 2 epochs of Tajik continual pre-training (merge-restacked per epoch).
- **Use:** a stronger Tajik base for generation and as the starting point for the next CPT epoch or for SFT.
- **Tajik MCQ (Zehnlab, full ~18,970 Q, loglikelihood):** see the eval repo `Tohirju/qwen35-tajik-eval-results`.
  For reference, base=44.7%, base+e1=48.9%.
- **Note:** this is a base/CPT model — it *completes* text, it does not follow chat instructions (that is the SFT stage).
- Merge verified by logit parity against the e2 adapter (max|Δlogit| < 1e0).
  Components: `unsloth/Qwen3.5-4B-Base` + `Tohirju/qwen35-4b-tajik-cpt-e1-merged` + `Tohirju/qwen35-4b-tajik-cpt-stacked-e2`.

Gated: download requires manual approval. Released by Saidzoda Lab / Saidzoda Engineering Company.