Text Generation
Transformers
Safetensors
Korean
llama
korean
causal-lm
decoder-only
from-scratch
instruction-tuned
1.2b
Eval Results (legacy)
text-generation-inference
Instructions to use properly59/Jumini-Ko-1.2B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use properly59/Jumini-Ko-1.2B with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="properly59/Jumini-Ko-1.2B")# Load model directly from transformers import AutoTokenizer, AutoModelForCausalLM tokenizer = AutoTokenizer.from_pretrained("properly59/Jumini-Ko-1.2B") model = AutoModelForCausalLM.from_pretrained("properly59/Jumini-Ko-1.2B", device_map="auto") - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use properly59/Jumini-Ko-1.2B with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "properly59/Jumini-Ko-1.2B" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "properly59/Jumini-Ko-1.2B", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'Use Docker
docker model run hf.co/properly59/Jumini-Ko-1.2B
- SGLang
How to use properly59/Jumini-Ko-1.2B with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "properly59/Jumini-Ko-1.2B" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "properly59/Jumini-Ko-1.2B", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "properly59/Jumini-Ko-1.2B" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "properly59/Jumini-Ko-1.2B", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }' - Docker Model Runner
How to use properly59/Jumini-Ko-1.2B with Docker Model Runner:
docker model run hf.co/properly59/Jumini-Ko-1.2B
Polish model card and add leaderboard status
Browse files
README.md
CHANGED
|
@@ -33,23 +33,25 @@ model-index:
|
|
| 33 |
|
| 34 |
# Jumini-Ko-1.2B
|
| 35 |
|
| 36 |
-
Jumini-Ko-1.2B is a Korean decoder-only language model trained from scratch for a university HW2 project.
|
| 37 |
|
| 38 |
-
This is a research checkpoint
|
| 39 |
|
| 40 |
-
##
|
| 41 |
|
| 42 |
| Field | Value |
|
| 43 |
|---|---|
|
|
|
|
|
|
|
| 44 |
| Parameters | 1,260,505,088 |
|
| 45 |
| Architecture | Decoder-only Transformer, RoPE, RMSNorm, SwiGLU, GQA |
|
| 46 |
| Layers / hidden size | 28 / 2048 |
|
| 47 |
| Vocabulary | 128,000 tokens |
|
| 48 |
-
| Export
|
| 49 |
| Primary language | Korean |
|
| 50 |
-
|
|
| 51 |
-
|
|
| 52 |
-
|
|
| 53 |
|
| 54 |
Source checkpoint:
|
| 55 |
|
|
@@ -57,7 +59,15 @@ Source checkpoint:
|
|
| 57 |
jumini-ko-1.2b-firsttok7000-argmax-boundary-layernormshock-v1-last16-4gpu0123-mb8-probe25-rerun20260608 / step_00007020
|
| 58 |
```
|
| 59 |
|
| 60 |
-
##
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 61 |
|
| 62 |
```python
|
| 63 |
import torch
|
|
@@ -84,17 +94,17 @@ outputs = model.generate(
|
|
| 84 |
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
|
| 85 |
```
|
| 86 |
|
| 87 |
-
The model does not currently ship with a polished chat template.
|
| 88 |
|
| 89 |
-
## Evaluation
|
| 90 |
|
| 91 |
-
The following
|
| 92 |
|
| 93 |
### Public Refresh Evidence
|
| 94 |
|
| 95 |
This upload replaced the earlier public `v3a6200` checkpoint because the broader local Korean sentinel improved from zero semantic hits to measurable pass@k behavior.
|
| 96 |
|
| 97 |
-
| Diagnostic | Previous HF `v3a6200` | Current `clean7020` |
|
| 98 |
|---|---:|---:|---:|
|
| 99 |
| Stratified Korean semantic pass@k | 0 / 240 = 0.0 | 21 / 240 = 8.8 | +8.8 |
|
| 100 |
| Stratified rollout pass rate | 0 / 2094 = 0.0 | 29 / 2105 = 1.4 | +1.4 |
|
|
@@ -102,7 +112,7 @@ This upload replaced the earlier public `v3a6200` checkpoint because the broader
|
|
| 102 |
|
| 103 |
### Jumini Score Appendix
|
| 104 |
|
| 105 |
-
The comparison table
|
| 106 |
|
| 107 |
Core clean7020 diagnostics:
|
| 108 |
|
|
@@ -130,11 +140,11 @@ Korean public/proxy stagegate scores:
|
|
| 130 |
| HRM8K subset | 2 / 60 = 3.3 |
|
| 131 |
| GSM8K-Ko subset | 0 / 100 = 0.0 |
|
| 132 |
|
| 133 |
-
Additional diagnostic scores in `jumini_internal_scores_20260609.json` include category breakdowns, strict heldout hit IDs, boundary-pair logit margins, and train-only generation forensics. The key boundary
|
| 134 |
|
| 135 |
### 1B-Class Public Baseline Comparison
|
| 136 |
|
| 137 |
-
For a size-adjacent open baseline, we compare against `TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T`, a widely used 1.1B English base model. This is
|
| 138 |
|
| 139 |
Both models were evaluated on the same 90-prompt Korean sentinel using greedy + 32 sampled rollouts per prompt (`temperature=0.7`, `top_p=0.95`, `top_k=50`, `max_new_tokens=64`).
|
| 140 |
|
|
@@ -152,7 +162,18 @@ Category breakdown:
|
|
| 152 |
|
| 153 |
Interpretation: Jumini does not dominate every subcategory, but it wins the overall Korean semantic pass@33 score and covers math/reasoning cases that the English-only base model misses.
|
| 154 |
|
| 155 |
-
##
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 156 |
|
| 157 |
| Check | Result |
|
| 158 |
|---|---|
|
|
@@ -161,7 +182,15 @@ Interpretation: Jumini does not dominate every subcategory, but it wins the over
|
|
| 161 |
| Argmax mismatch rate vs native checkpoint | 0.0 |
|
| 162 |
| Max absolute logit diff | 1.478e-5 |
|
| 163 |
|
| 164 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 165 |
|
| 166 |
## Intended Use
|
| 167 |
|
|
@@ -170,32 +199,30 @@ This model is intended for:
|
|
| 170 |
- Korean language modeling and generation experiments.
|
| 171 |
- Course/project reproducibility.
|
| 172 |
- Studying from-scratch small-LLM training behavior under limited compute.
|
| 173 |
-
- Diagnostic comparison against
|
| 174 |
|
| 175 |
It is not intended for:
|
| 176 |
|
| 177 |
- Safety-critical, medical, legal, or financial advice.
|
| 178 |
- Production deployment without additional safety evaluation.
|
| 179 |
-
- Claims of official leaderboard performance
|
| 180 |
|
| 181 |
## Limitations
|
| 182 |
|
| 183 |
-
- Greedy decoding remains weak on the internal Korean semantic sentinel (`0/90` greedy hits in the K32 run).
|
| 184 |
-
- The strict heldout project gate remains limited at `2/12`; this refresh is a best-so-far public update, not a final breakthrough checkpoint.
|
| 185 |
- The model can repeat, terminate poorly, or produce malformed outputs.
|
| 186 |
- The comparison above is a Korean local diagnostic. It does not imply superiority over strong Korean-specialized models such as EXAONE-4.0-1.2B.
|
| 187 |
- The model has not undergone full safety alignment or red-team evaluation.
|
| 188 |
|
| 189 |
## Training and Provenance
|
| 190 |
|
| 191 |
-
The model implementation, tokenizer pipeline, data filtering, training loop, checkpointing, evaluation scripts, and HF export scripts were implemented within the HW2 project. The released checkpoint is not a fine-tune of TinyLlama, Llama, EXAONE, Qwen, or another public model.
|
| 192 |
|
| 193 |
Project reports and manifests document data provenance, filtering, and evaluation decisions. The public artifact is provided under Apache-2.0.
|
| 194 |
|
| 195 |
## Citation
|
| 196 |
|
| 197 |
-
If you reference this model, please cite the repository and checkpoint name:
|
| 198 |
-
|
| 199 |
```text
|
| 200 |
Jumini-Ko-1.2B clean7020, properly59/Jumini-Ko-1.2B, 2026-06-09 public refresh.
|
| 201 |
```
|
|
|
|
| 33 |
|
| 34 |
# Jumini-Ko-1.2B
|
| 35 |
|
| 36 |
+
Jumini-Ko-1.2B is a Korean decoder-only language model trained from scratch for a university HW2 project. This repository hosts the current best public refresh as of 2026-06-09, replacing an earlier public checkpoint after measurable gains on local Korean generation diagnostics.
|
| 37 |
|
| 38 |
+
This is a research checkpoint. It is released for inspection, reproduction, and course evaluation, with both strengths and failure modes documented.
|
| 39 |
|
| 40 |
+
## Release Snapshot
|
| 41 |
|
| 42 |
| Field | Value |
|
| 43 |
|---|---|
|
| 44 |
+
| Model | `properly59/Jumini-Ko-1.2B` |
|
| 45 |
+
| Public checkpoint | `clean7020` |
|
| 46 |
| Parameters | 1,260,505,088 |
|
| 47 |
| Architecture | Decoder-only Transformer, RoPE, RMSNorm, SwiGLU, GQA |
|
| 48 |
| Layers / hidden size | 28 / 2048 |
|
| 49 |
| Vocabulary | 128,000 tokens |
|
| 50 |
+
| Export format | `LlamaForCausalLM`-compatible Transformers + Safetensors |
|
| 51 |
| Primary language | Korean |
|
| 52 |
+
| License | Apache-2.0 |
|
| 53 |
+
| Release date | 2026-06-09 |
|
| 54 |
+
| HF revision for current score appendix | `aec6d67bf3ce37e5845e0fcd8f05ad1deb322f50` |
|
| 55 |
|
| 56 |
Source checkpoint:
|
| 57 |
|
|
|
|
| 59 |
jumini-ko-1.2b-firsttok7000-argmax-boundary-layernormshock-v1-last16-4gpu0123-mb8-probe25-rerun20260608 / step_00007020
|
| 60 |
```
|
| 61 |
|
| 62 |
+
## Highlights
|
| 63 |
+
|
| 64 |
+
- From-scratch Korean 1.2B-class decoder-only model; not a fine-tune of Llama, Qwen, EXAONE, TinyLlama, or Gemma.
|
| 65 |
+
- Replaced the original public HF upload after broad local Korean semantic pass@k improved from `0 / 240` to `21 / 240`.
|
| 66 |
+
- Wins a size-adjacent local Korean specialization comparison against `TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T` on Korean semantic pass@33.
|
| 67 |
+
- Ships with machine-readable validation, benchmark, replacement, and score appendix files.
|
| 68 |
+
- Known weak point: greedy decoding remains unreliable, and strict heldout performance is still limited.
|
| 69 |
+
|
| 70 |
+
## Quickstart
|
| 71 |
|
| 72 |
```python
|
| 73 |
import torch
|
|
|
|
| 94 |
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
|
| 95 |
```
|
| 96 |
|
| 97 |
+
Plain Korean prompts are recommended. The model does not currently ship with a polished chat template. Sampling is usually more useful than greedy decoding for this checkpoint.
|
| 98 |
|
| 99 |
+
## Evaluation Summary
|
| 100 |
|
| 101 |
+
The following numbers are local project diagnostics, not official leaderboard scores. They use deterministic rule-based verifiers on Korean sentinel prompts and public/proxy Korean multiple-choice subsets. No external teacher model or model-judge score is used for these local diagnostics.
|
| 102 |
|
| 103 |
### Public Refresh Evidence
|
| 104 |
|
| 105 |
This upload replaced the earlier public `v3a6200` checkpoint because the broader local Korean sentinel improved from zero semantic hits to measurable pass@k behavior.
|
| 106 |
|
| 107 |
+
| Diagnostic | Previous HF `v3a6200` | Current `clean7020` | Delta |
|
| 108 |
|---|---:|---:|---:|
|
| 109 |
| Stratified Korean semantic pass@k | 0 / 240 = 0.0 | 21 / 240 = 8.8 | +8.8 |
|
| 110 |
| Stratified rollout pass rate | 0 / 2094 = 0.0 | 29 / 2105 = 1.4 | +1.4 |
|
|
|
|
| 112 |
|
| 113 |
### Jumini Score Appendix
|
| 114 |
|
| 115 |
+
The compact comparison table is not the full score record. A fuller Jumini-only score appendix for the current public `clean7020` checkpoint is included in `jumini_internal_scores_20260609.json`. It contains completed score summaries while omitting row-level prompt text, completions, and non-exported descendant checkpoints.
|
| 116 |
|
| 117 |
Core clean7020 diagnostics:
|
| 118 |
|
|
|
|
| 140 |
| HRM8K subset | 2 / 60 = 3.3 |
|
| 141 |
| GSM8K-Ko subset | 0 / 100 = 0.0 |
|
| 142 |
|
| 143 |
+
Additional diagnostic scores in `jumini_internal_scores_20260609.json` include category breakdowns, strict heldout hit IDs, boundary-pair logit margins, and train-only generation forensics. The key boundary diagnostic was `327 / 418 = 78.2` target-in-top-k on accepted pass@k boundary pairs, but only `1 / 418 = 0.2` target margin >= 0. This matches the observed weakness that useful sampled behavior does not reliably become greedy behavior.
|
| 144 |
|
| 145 |
### 1B-Class Public Baseline Comparison
|
| 146 |
|
| 147 |
+
For a size-adjacent open baseline, we compare against `TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T`, a widely used 1.1B English base model. This is a Korean-specialization diagnostic: it should be read as "Jumini is better on this Korean local sentinel", not as a claim that Jumini is globally stronger.
|
| 148 |
|
| 149 |
Both models were evaluated on the same 90-prompt Korean sentinel using greedy + 32 sampled rollouts per prompt (`temperature=0.7`, `top_p=0.95`, `top_k=50`, `max_new_tokens=64`).
|
| 150 |
|
|
|
|
| 162 |
|
| 163 |
Interpretation: Jumini does not dominate every subcategory, but it wins the overall Korean semantic pass@33 score and covers math/reasoning cases that the English-only base model misses.
|
| 164 |
|
| 165 |
+
## Leaderboard Submission Status
|
| 166 |
+
|
| 167 |
+
Leaderboard requests have been submitted through public Hugging Face requests-dataset PR flows. They are pending maintainer merge/evaluation; no official leaderboard score is claimed here before completion.
|
| 168 |
+
|
| 169 |
+
| Leaderboard | Request | Status checked | Revision | Precision | Model type |
|
| 170 |
+
|---|---|---|---|---|---|
|
| 171 |
+
| Open Ko-LLM Leaderboard | `https://huggingface.co/datasets/open-ko-llm-leaderboard/requests/discussions/2` | `open`, pending maintainer merge/evaluation | `aec6d67bf3ce37e5845e0fcd8f05ad1deb322f50` | `float16` | `pretrained model` / `Original` |
|
| 172 |
+
| Open LLM Leaderboard | `https://huggingface.co/datasets/open-llm-leaderboard/requests/discussions/92` | `open`, pending maintainer merge/evaluation | `aec6d67bf3ce37e5845e0fcd8f05ad1deb322f50` | `float16` | `pretrained` / `Original` |
|
| 173 |
+
|
| 174 |
+
The submitted revision is pinned to the model-card and score-appendix refresh in this repository. The Open Ko request commit is `942fc56e6fdb68cc80658527e3f96e98c92f21f9`; the Open LLM request commit is `a570c5fc07c763ef4b06995870066d14c440a87e`.
|
| 175 |
+
|
| 176 |
+
## Validation Artifacts
|
| 177 |
|
| 178 |
| Check | Result |
|
| 179 |
|---|---|
|
|
|
|
| 182 |
| Argmax mismatch rate vs native checkpoint | 0.0 |
|
| 183 |
| Max absolute logit diff | 1.478e-5 |
|
| 184 |
|
| 185 |
+
Machine-readable artifacts:
|
| 186 |
+
|
| 187 |
+
| File | Purpose |
|
| 188 |
+
|---|---|
|
| 189 |
+
| `load_smoke.json` | Transformers load and parameter smoke check |
|
| 190 |
+
| `hf_equivalence_atol2e-5.json` | Native-to-HF export equivalence |
|
| 191 |
+
| `replacement_evidence_20260609.json` | Previous HF checkpoint vs clean7020 replacement evidence |
|
| 192 |
+
| `benchmark_results_20260609.json` | Public comparison summary |
|
| 193 |
+
| `jumini_internal_scores_20260609.json` | Fuller clean7020 score appendix |
|
| 194 |
|
| 195 |
## Intended Use
|
| 196 |
|
|
|
|
| 199 |
- Korean language modeling and generation experiments.
|
| 200 |
- Course/project reproducibility.
|
| 201 |
- Studying from-scratch small-LLM training behavior under limited compute.
|
| 202 |
+
- Diagnostic comparison against compact open models on Korean prompts.
|
| 203 |
|
| 204 |
It is not intended for:
|
| 205 |
|
| 206 |
- Safety-critical, medical, legal, or financial advice.
|
| 207 |
- Production deployment without additional safety evaluation.
|
| 208 |
+
- Claims of official leaderboard performance before the relevant public benchmark has completed.
|
| 209 |
|
| 210 |
## Limitations
|
| 211 |
|
| 212 |
+
- Greedy decoding remains weak on the internal Korean semantic sentinel (`0 / 90` greedy hits in the K32 run).
|
| 213 |
+
- The strict heldout project gate remains limited at `2 / 12`; this refresh is a best-so-far public update, not a final breakthrough checkpoint.
|
| 214 |
- The model can repeat, terminate poorly, or produce malformed outputs.
|
| 215 |
- The comparison above is a Korean local diagnostic. It does not imply superiority over strong Korean-specialized models such as EXAONE-4.0-1.2B.
|
| 216 |
- The model has not undergone full safety alignment or red-team evaluation.
|
| 217 |
|
| 218 |
## Training and Provenance
|
| 219 |
|
| 220 |
+
The model implementation, tokenizer pipeline, data filtering, training loop, checkpointing, evaluation scripts, and HF export scripts were implemented within the HW2 project. The released checkpoint is not a fine-tune of TinyLlama, Llama, EXAONE, Qwen, Gemma, or another public model.
|
| 221 |
|
| 222 |
Project reports and manifests document data provenance, filtering, and evaluation decisions. The public artifact is provided under Apache-2.0.
|
| 223 |
|
| 224 |
## Citation
|
| 225 |
|
|
|
|
|
|
|
| 226 |
```text
|
| 227 |
Jumini-Ko-1.2B clean7020, properly59/Jumini-Ko-1.2B, 2026-06-09 public refresh.
|
| 228 |
```
|