v0.2
Browse files- README.md +11 -10
- config.json +7 -7
- model.safetensors +2 -2
README.md
CHANGED
|
@@ -18,30 +18,31 @@ model-index:
|
|
| 18 |
|
| 19 |
**The smallest coherent Hungarian language model.** Part of the [Emese](https://emese.tech) model family.
|
| 20 |
|
| 21 |
-
Csermely is a
|
| 22 |
|
| 23 |
## Model Details
|
| 24 |
|
| 25 |
| | |
|
| 26 |
|---|---|
|
| 27 |
-
| **
|
| 28 |
-
| **
|
|
|
|
| 29 |
| **Architecture** | LLaMA-style (decoder-only transformer) |
|
| 30 |
-
| **Training context** |
|
| 31 |
| **Training precision** | bfloat16 (MLX) |
|
| 32 |
| **Published weights** | float16 |
|
| 33 |
| **Vocabulary** | 32,000 (SentencePiece Unigram, Hungarian) |
|
| 34 |
-
| **Training data** | ~
|
| 35 |
| **License** | MIT |
|
| 36 |
|
| 37 |
## Architecture
|
| 38 |
|
| 39 |
- 16 transformer layers
|
| 40 |
-
-
|
| 41 |
-
-
|
| 42 |
-
-
|
| 43 |
- RMSNorm pre-layer normalization
|
| 44 |
-
- Rotary positional embeddings (RoPE) with YaRN extension
|
| 45 |
- SwiGLU feed-forward activation
|
| 46 |
- Tied input/output embeddings
|
| 47 |
|
|
@@ -71,7 +72,7 @@ The default generation config uses `temperature=0.7`, `top_p=0.9`, and `repetiti
|
|
| 71 |
|
| 72 |
```bibtex
|
| 73 |
@misc{emese-csermely-2026,
|
| 74 |
-
title={Csermely: A
|
| 75 |
author={Emese Tech},
|
| 76 |
year={2026},
|
| 77 |
url={https://huggingface.co/emese-tech/csermely}
|
|
|
|
| 18 |
|
| 19 |
**The smallest coherent Hungarian language model.** Part of the [Emese](https://emese.tech) model family.
|
| 20 |
|
| 21 |
+
Csermely is a 190M parameter decoder-only transformer trained exclusively on high-quality Hungarian text. It runs on edge devices and excels in summarization, grammar checking, and tone detection.
|
| 22 |
|
| 23 |
## Model Details
|
| 24 |
|
| 25 |
| | |
|
| 26 |
|---|---|
|
| 27 |
+
| **Version** | 0.2 |
|
| 28 |
+
| **Parameters** | 190.2M |
|
| 29 |
+
| **Context length** | 4,096 tokens (YaRN RoPE, 4× factor) |
|
| 30 |
| **Architecture** | LLaMA-style (decoder-only transformer) |
|
| 31 |
+
| **Training context** | 1,024 tokens |
|
| 32 |
| **Training precision** | bfloat16 (MLX) |
|
| 33 |
| **Published weights** | float16 |
|
| 34 |
| **Vocabulary** | 32,000 (SentencePiece Unigram, Hungarian) |
|
| 35 |
+
| **Training data** | ~2B tokens of Hungarian text |
|
| 36 |
| **License** | MIT |
|
| 37 |
|
| 38 |
## Architecture
|
| 39 |
|
| 40 |
- 16 transformer layers
|
| 41 |
+
- 896 hidden dimension
|
| 42 |
+
- 14 attention heads
|
| 43 |
+
- 2560 FFN intermediate size
|
| 44 |
- RMSNorm pre-layer normalization
|
| 45 |
+
- Rotary positional embeddings (RoPE) with YaRN extension (4× factor, base 1024)
|
| 46 |
- SwiGLU feed-forward activation
|
| 47 |
- Tied input/output embeddings
|
| 48 |
|
|
|
|
| 72 |
|
| 73 |
```bibtex
|
| 74 |
@misc{emese-csermely-2026,
|
| 75 |
+
title={Csermely: A Hungarian Language Model},
|
| 76 |
author={Emese Tech},
|
| 77 |
year={2026},
|
| 78 |
url={https://huggingface.co/emese-tech/csermely}
|
config.json
CHANGED
|
@@ -1,20 +1,20 @@
|
|
| 1 |
{
|
| 2 |
"architectures": ["LlamaForCausalLM"],
|
| 3 |
"model_type": "llama",
|
| 4 |
-
"model_version": "0.
|
| 5 |
"vocab_size": 32000,
|
| 6 |
-
"hidden_size":
|
| 7 |
-
"intermediate_size":
|
| 8 |
"num_hidden_layers": 16,
|
| 9 |
-
"num_attention_heads":
|
| 10 |
-
"num_key_value_heads":
|
| 11 |
"hidden_act": "silu",
|
| 12 |
-
"max_position_embeddings":
|
| 13 |
"rope_theta": 10000.0,
|
| 14 |
"rope_scaling": {
|
| 15 |
"type": "yarn",
|
| 16 |
"factor": 4.0,
|
| 17 |
-
"original_max_position_embeddings":
|
| 18 |
},
|
| 19 |
"rms_norm_eps": 1e-5,
|
| 20 |
"tie_word_embeddings": true,
|
|
|
|
| 1 |
{
|
| 2 |
"architectures": ["LlamaForCausalLM"],
|
| 3 |
"model_type": "llama",
|
| 4 |
+
"model_version": "0.2",
|
| 5 |
"vocab_size": 32000,
|
| 6 |
+
"hidden_size": 896,
|
| 7 |
+
"intermediate_size": 2560,
|
| 8 |
"num_hidden_layers": 16,
|
| 9 |
+
"num_attention_heads": 14,
|
| 10 |
+
"num_key_value_heads": 14,
|
| 11 |
"hidden_act": "silu",
|
| 12 |
+
"max_position_embeddings": 4096,
|
| 13 |
"rope_theta": 10000.0,
|
| 14 |
"rope_scaling": {
|
| 15 |
"type": "yarn",
|
| 16 |
"factor": 4.0,
|
| 17 |
+
"original_max_position_embeddings": 1024
|
| 18 |
},
|
| 19 |
"rms_norm_eps": 1e-5,
|
| 20 |
"tie_word_embeddings": true,
|
model.safetensors
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:deeac4d68a23cc62416245c42d172caaefb1db6b324d50027e7ea98db045e5f1
|
| 3 |
+
size 437724864
|