HyperCLOVA-X-SEED-Think-14B Minimal SFT KRC MONO v/o LoRA
This is a merged Korean causal-language-model checkpoint trained for the KDS task. It is derived from the local hcx_minimal_sft_71875 checkpoint, whose upstream base is HyperCLOVA-X-SEED-Think-14B.
Training
- Starting checkpoint:
models/hcx_minimal_sft_71875 - Training data: 836 fixed K/R/C success-consolidation MONO examples (
MONO_train.jsonl, SHA-256f5c98fc8366ca99c4d83fa723a521108fa1d266a9a8b15c66b404da545e42023) - Method: LoRA merged into the base checkpoint
- LoRA configuration: rank 4, alpha 8, dropout 0.05; target modules
v_projando_proj - Learning rate: 5e-6, cosine schedule, warmup ratio 0.03
- One epoch, effective batch size 16 (per-device batch size 4, gradient accumulation 4), BF16
- Trainable parameters: 2,957,312
The data was reused as the existing 836-example consolidated set; K/R/C was not re-solved or re-sampled for this run.
Run and validation
- Run ID:
TRAIN-20260908-HCX-MINIMAL-SFT-KRC-MONO-VO-LR5E-6-GPU0 - Final training loss:
0.5008622160497701 - Training completed: 2026-09-08 16:28:18 UTC
- Merge completed on a single GPU: 2026-09-08 16:31:27 UTC
- The merged checkpoint contains 16 BF16 SafeTensors shards (about 27.47 GiB).
- Adapter and merged-checkpoint NaN/Inf checks passed.
- Generation smoke tests passed for arithmetic and Korean social-situation prompts. The chat template is unchanged from the training base.
Local benchmark and K-AI leaderboard evaluation were not run for this checkpoint. The loss and smoke-test results should not be interpreted as a leaderboard score.
Usage
Use this model with a recent transformers release and the included tokenizer and chat template. It is a merged model checkpoint, so no separate PEFT adapter is required.
License and notices
Refer to the bundled LICENSE and NOTICE files. Use is subject to the applicable license and terms for the upstream model.