properly59 commited on
Commit
2909894
·
verified ·
1 Parent(s): aec6d67

Polish model card and add leaderboard status

Browse files
Files changed (1) hide show
  1. README.md +51 -24
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. The checkpoint in this repository is the current best public refresh as of 2026-06-09: it replaces an earlier public upload after improving broad local Korean generation diagnostics.
37
 
38
- This is a research checkpoint, not a production assistant. It is released to make the training artifact inspectable and reproducible, including known limitations and diagnostic results.
39
 
40
- ## Model Details
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 class | `LlamaForCausalLM`-compatible Transformers layout |
49
  | Primary language | Korean |
50
- | Training status | From-scratch pretraining + project SFT/probe stages |
51
- | Current public checkpoint | `clean7020` best-so-far refresh |
52
- | Replacement date | 2026-06-09 |
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
- ## How to Use
 
 
 
 
 
 
 
 
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. Plain Korean prompts are recommended. Sampling is often more useful than greedy decoding for this checkpoint.
88
 
89
- ## Evaluation
90
 
91
- The following results are local project diagnostics, not official leaderboard scores. They use deterministic rule-based verifiers on train-only Korean semantic sentinel prompts. No external teacher, model judge, or heldout prompt text is used for these diagnostics.
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` | Score Delta |
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 below is intentionally compact. The fuller Jumini-only score record for the current public `clean7020` checkpoint is included in `jumini_internal_scores_20260609.json`. It includes completed score summaries for the current public model while omitting row-level prompt text, completions, and non-exported descendant checkpoints.
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 diagnostics were `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.
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 intentionally 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.
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
- ### Validation Artifacts
 
 
 
 
 
 
 
 
 
 
 
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
- See `load_smoke.json`, `hf_equivalence_atol2e-5.json`, `replacement_evidence_20260609.json`, `benchmark_results_20260609.json`, and `jumini_internal_scores_20260609.json` in this repository for machine-readable details.
 
 
 
 
 
 
 
 
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 other compact open models on Korean prompts.
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 without running the relevant public benchmark suite.
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
  ```