Jamba2 3B Turkish SFT v1
This is the BF16 merged instruction-tuning artifact produced from the immutable
Turkish CPT base serda-dev/Jamba2-3B-Turkish revision
9524af9e857e0359b8a3dad72bc216b65d3c0acd.
Training
- Method: BF16 LoRA, merged into the base model.
- Selected checkpoint: optimizer step 4,500 (not the final step).
- SFT mixture: 35,800 reviewed/generated Turkish instruction examples.
- Training run: 5,000 optimizer steps, 40,960,000 tokens.
Diagnostic evaluation
The isolated 84-case Gold suite is draft_quarantine, so these results are
diagnostic rather than release-grade. The post-merge deterministic evaluation
reproduced the selected checkpoint exactly:
| Metric | Result |
|---|---|
| Semantic/task correctness | 62/84 |
| Contract compliance | 58/84 |
gold.json, gold.md, merge_manifest.json, and SHA256SUMS accompany the
weights for provenance and integrity checks. Final production promotion still
requires the Gold-freeze, overlap/decontamination, and human-review gates.
Usage
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "linguai/Jamba2-3B-Turkish-SFT-v1"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype=torch.bfloat16,
device_map="auto",
)
Use the model according to the terms and conditions of the upstream base model and all included training sources.