pravdin commited on
Commit
b81105f
·
verified ·
1 Parent(s): a1b110b

Upload folder using huggingface_hub

Browse files
README.md CHANGED
@@ -27,8 +27,8 @@ This model represents a **systematic exploration** of the combination of reasoni
27
  ## 🔬 Model Lineage & Methodology
28
 
29
  ### Parent Models
30
- - **Primary**: [DeepSeek-R1-Distill-Qwen-1.5B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B) - A model trained via large-scale reinforcement learning, demonstrating advanced reasoning capabilities across various benchmarks.
31
- - **Secondary**: [Gensyn/Qwen2.5-1.5B-Instruct](https://huggingface.co/Gensyn/Qwen2.5-1.5B-Instruct) - A model designed for instruction-following tasks, showcasing proficiency in understanding and executing user prompts.
32
 
33
  ### Merge Configuration
34
  ```yaml
@@ -46,40 +46,40 @@ dtype: bfloat16
46
  ```
47
 
48
  ### Research Rationale
49
- The combination of DeepSeek-R1's reasoning capabilities with Qwen2.5's instruction-following abilities was hypothesized to create a model that excels in tasks requiring both logical reasoning and adherence to user instructions, thereby addressing the limitations observed in each individual model.
50
 
51
  ## 🎯 Intended Use & Research Applications
52
 
53
  ### Primary Research Use Cases
54
  - Complex problem-solving in mathematics and reasoning tasks
55
  - Instruction-following applications in educational tools
56
- - Benchmarking performance in natural language understanding and generation
57
 
58
  ### Production Considerations
59
- While this model shows promise in enhancing reasoning and instruction-following tasks, it is essential to consider the potential for overfitting to specific types of queries and the need for careful prompt engineering to achieve optimal results.
60
 
61
  ## 📊 Evaluation & Validation
62
 
63
  ### Research Metrics
64
- The model's performance was evaluated using a range of benchmarks, including MMLU, DROP, and LiveCodeBench, with results indicating improved performance in reasoning tasks compared to baseline models.
65
 
66
  ### Known Capabilities
67
  - Enhanced reasoning patterns through the integration of reinforcement learning techniques
68
- - Improved instruction adherence in complex queries
69
 
70
  ### Performance Characteristics
71
- Quantitative results indicate that the model outperforms individual parent models in several benchmarks, particularly in tasks requiring both reasoning and instruction adherence.
72
 
73
  ## ⚠️ Limitations & Research Boundaries
74
 
75
  ### Technical Limitations
76
- The model may exhibit limitations in generating coherent outputs for highly complex queries, particularly those that require extensive context or multi-step reasoning.
77
 
78
  ### Research Scope
79
- This research does not explore the full range of capabilities of either parent model but focuses specifically on the intersection of reasoning and instruction-following.
80
 
81
  ### Ethical Considerations
82
- Potential biases inherent in the training data of the parent models may affect the outputs of the merged model. Responsible use guidelines should be followed to mitigate risks associated with biased outputs.
83
 
84
  ## 🔬 Research Framework
85
 
 
27
  ## 🔬 Model Lineage & Methodology
28
 
29
  ### Parent Models
30
+ - **Primary**: [DeepSeek-R1-Distill-Qwen-1.5B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B) - A distilled model that demonstrates advanced reasoning capabilities through reinforcement learning, achieving competitive performance across various benchmarks.
31
+ - **Secondary**: [Gensyn/Qwen2.5-1.5B-Instruct](https://huggingface.co/Gensyn/Qwen2.5-1.5B-Instruct) - A model designed for instruction-following tasks, providing a strong foundation for generating coherent and contextually relevant responses.
32
 
33
  ### Merge Configuration
34
  ```yaml
 
46
  ```
47
 
48
  ### Research Rationale
49
+ The combination of DeepSeek-R1's reasoning capabilities with Qwen2.5's instruction-following abilities was hypothesized to create a model that excels in tasks requiring both logical reasoning and adherence to user prompts, thereby addressing limitations observed in individual models.
50
 
51
  ## 🎯 Intended Use & Research Applications
52
 
53
  ### Primary Research Use Cases
54
  - Complex problem-solving in mathematics and reasoning tasks
55
  - Instruction-following applications in educational tools
56
+ - Benchmarking against state-of-the-art models in reasoning and language understanding
57
 
58
  ### Production Considerations
59
+ While this model shows promise in enhancing reasoning and instruction-following tasks, it is essential to consider the potential for overfitting to specific types of queries and the need for careful prompt engineering to achieve optimal performance.
60
 
61
  ## 📊 Evaluation & Validation
62
 
63
  ### Research Metrics
64
+ Evaluation was conducted using a variety of benchmarks, including MMLU, DROP, and LiveCodeBench, with results indicating improved performance in reasoning tasks compared to baseline models.
65
 
66
  ### Known Capabilities
67
  - Enhanced reasoning patterns through the integration of reinforcement learning techniques
68
+ - Improved instruction adherence and contextual understanding in generated responses
69
 
70
  ### Performance Characteristics
71
+ Quantitative results indicate that the model outperforms individual parent models in several key benchmarks, demonstrating the effectiveness of the merging strategy.
72
 
73
  ## ⚠️ Limitations & Research Boundaries
74
 
75
  ### Technical Limitations
76
+ The model may exhibit limitations in handling highly specialized queries or tasks that require domain-specific knowledge beyond the training data.
77
 
78
  ### Research Scope
79
+ This research focuses on the integration of reasoning and instruction-following capabilities and does not explore other potential model combinations or architectures.
80
 
81
  ### Ethical Considerations
82
+ Bias risks inherent in the training data of the parent models may persist in the merged model. Responsible use guidelines should be followed to mitigate potential misuse.
83
 
84
  ## 🔬 Research Framework
85
 
config.json CHANGED
@@ -1,5 +1,4 @@
1
  {
2
- "_name_or_path": "Gensyn/Qwen2.5-1.5B-Instruct",
3
  "architectures": [
4
  "Qwen2ForCausalLM"
5
  ],
@@ -10,6 +9,36 @@
10
  "hidden_size": 1536,
11
  "initializer_range": 0.02,
12
  "intermediate_size": 8960,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
13
  "max_position_embeddings": 32768,
14
  "max_window_layers": 21,
15
  "model_type": "qwen2",
@@ -17,11 +46,12 @@
17
  "num_hidden_layers": 28,
18
  "num_key_value_heads": 2,
19
  "rms_norm_eps": 1e-06,
 
20
  "rope_theta": 1000000.0,
21
  "sliding_window": null,
22
  "tie_word_embeddings": true,
23
  "torch_dtype": "bfloat16",
24
- "transformers_version": "4.44.0",
25
  "use_cache": true,
26
  "use_sliding_window": false,
27
  "vocab_size": 151936
 
1
  {
 
2
  "architectures": [
3
  "Qwen2ForCausalLM"
4
  ],
 
9
  "hidden_size": 1536,
10
  "initializer_range": 0.02,
11
  "intermediate_size": 8960,
12
+ "layer_types": [
13
+ "full_attention",
14
+ "full_attention",
15
+ "full_attention",
16
+ "full_attention",
17
+ "full_attention",
18
+ "full_attention",
19
+ "full_attention",
20
+ "full_attention",
21
+ "full_attention",
22
+ "full_attention",
23
+ "full_attention",
24
+ "full_attention",
25
+ "full_attention",
26
+ "full_attention",
27
+ "full_attention",
28
+ "full_attention",
29
+ "full_attention",
30
+ "full_attention",
31
+ "full_attention",
32
+ "full_attention",
33
+ "full_attention",
34
+ "full_attention",
35
+ "full_attention",
36
+ "full_attention",
37
+ "full_attention",
38
+ "full_attention",
39
+ "full_attention",
40
+ "full_attention"
41
+ ],
42
  "max_position_embeddings": 32768,
43
  "max_window_layers": 21,
44
  "model_type": "qwen2",
 
46
  "num_hidden_layers": 28,
47
  "num_key_value_heads": 2,
48
  "rms_norm_eps": 1e-06,
49
+ "rope_scaling": null,
50
  "rope_theta": 1000000.0,
51
  "sliding_window": null,
52
  "tie_word_embeddings": true,
53
  "torch_dtype": "bfloat16",
54
+ "transformers_version": "4.53.0",
55
  "use_cache": true,
56
  "use_sliding_window": false,
57
  "vocab_size": 151936
model-00001-of-00004.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:8fbeffabb36c64e581959ebc641c0d09961cf38c8319ff1069fdd14b471cd569
3
  size 988548664
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f65df8ba299179ed078158f700fd523d8dd179cb68de306a0da52ebb33496ff6
3
  size 988548664
model-00002-of-00004.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:f64d563aa0f608e1312ea5efdd6a03e0d4b0587d3aa0bc553d8ca2be3f3e2af4
3
  size 974516536
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b57bf8aba2b675dc1391a187b383343e8fbd7c44e50203d0c2215c3f5a3fdde3
3
  size 974516536
model-00003-of-00004.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:34a16b908c0a31d6f997ba5534af6c7be0dd4db3aaa9007fc36a1692cec0f184
3
  size 991020472
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b2c1830719b56e9e836c9165b5175aff5e76225b17db98e6634a03623eb202b3
3
  size 991020472
model-00004-of-00004.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:b8267a0db5ca06da83f5dc5c1a4da2543766734e33b29f47b6b620ee0722a7df
3
  size 600128448
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d9ab763cc47a0fbf8c72e2f7ab5a7abb76c802df19131f44449612a538669210
3
  size 600128448