|
|
--- |
|
|
license: apache-2.0 |
|
|
language: |
|
|
- en |
|
|
base_model: |
|
|
- deepseek-ai/DeepSeek-R1-Distill-Qwen-14B |
|
|
pipeline_tag: text-generation |
|
|
library_name: transformers |
|
|
tags: |
|
|
- text-generation-inference |
|
|
- math |
|
|
- moderately abliterated |
|
|
- abliterated |
|
|
- code |
|
|
- R1 |
|
|
- RL |
|
|
--- |
|
|
|
|
|
 |
|
|
|
|
|
# **Sombrero-R1-14B-Elite13** |
|
|
|
|
|
> Sombrero-R1-14B-Elite13 is a fine-tuned variant of the DeepSeek-R1-Distill-Qwen-14B model, enhanced through reinforcement learning to serve as a high-performance reasoning assistant. It excels in both mathematical problem-solving and general-purpose conversational tasks. This model combines distilled efficiency with refined instruction-following behavior, offering an ideal balance of speed, capability, and coherence for complex interactive tasks. |
|
|
|
|
|
### Key Enhancements |
|
|
|
|
|
1. **Reinforcement Learning Fine-Tuning** |
|
|
Trained with reinforcement learning objectives to optimize for alignment, reward-guided reasoning, and helpfulness in conversation. |
|
|
|
|
|
2. **Mathematical Reasoning Proficiency** |
|
|
Delivers accurate solutions and step-by-step breakdowns for algebra, calculus, number theory, logic puzzles, and applied mathematics. |
|
|
|
|
|
3. **Instruction Adherence** |
|
|
Capable of understanding and following multi-part instructions, including structured tasks and iterative refinement prompts. |
|
|
|
|
|
4. **Expanded Context Handling** |
|
|
Supports up to 128K tokens of context with output lengths up to 8K tokens, ideal for technical and educational use cases. |
|
|
|
|
|
5. **Cross-Domain Knowledge** |
|
|
Offers broad general knowledge capabilities, making it suitable for tutoring, research, and exploratory conversation across topics. |
|
|
|
|
|
--- |
|
|
|
|
|
# **Quickstart with Transformers** |
|
|
|
|
|
```python |
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer |
|
|
|
|
|
model_name = "prithivMLmods/Sombrero-R1-14B-Elite13" |
|
|
|
|
|
model = AutoModelForCausalLM.from_pretrained( |
|
|
model_name, |
|
|
torch_dtype="auto", |
|
|
device_map="auto" |
|
|
) |
|
|
tokenizer = AutoTokenizer.from_pretrained(model_name) |
|
|
|
|
|
prompt = "Solve: Integrate (x^2 * e^x) dx" |
|
|
messages = [ |
|
|
{"role": "system", "content": "You are a helpful AI assistant skilled in math and reasoning."}, |
|
|
{"role": "user", "content": prompt} |
|
|
] |
|
|
text = tokenizer.apply_chat_template( |
|
|
messages, |
|
|
tokenize=False, |
|
|
add_generation_prompt=True |
|
|
) |
|
|
model_inputs = tokenizer([text], return_tensors="pt").to(model.device) |
|
|
|
|
|
generated_ids = model.generate( |
|
|
**model_inputs, |
|
|
max_new_tokens=512 |
|
|
) |
|
|
generated_ids = [ |
|
|
output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids) |
|
|
] |
|
|
|
|
|
response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0] |
|
|
``` |
|
|
|
|
|
--- |
|
|
|
|
|
# **Intended Use Cases** |
|
|
|
|
|
1. **Mathematics Problem Solving** |
|
|
Ideal for step-by-step derivations, symbolic computation, numerical explanations, and LaTeX-supported outputs. |
|
|
|
|
|
2. **Educational and Instructional Support** |
|
|
Helpful in classrooms and learning platforms, offering guided explanations for students and instructors. |
|
|
|
|
|
3. **Chat-based Reasoning** |
|
|
Designed for coherent, context-aware dialogue generation with structured logic and continuity. |
|
|
|
|
|
4. **Multilingual Knowledge Assistance** |
|
|
Supports 29+ languages, including English, Chinese, French, German, Arabic, and others, for multilingual learning. |
|
|
|
|
|
5. **Document and Code Explanation** |
|
|
Can explain complex documents, code snippets, or structured logic flows in natural language. |
|
|
|
|
|
--- |
|
|
|
|
|
# **Known Limitations** |
|
|
|
|
|
1. **Compute Intensive** |
|
|
Requires high-memory hardware (e.g., ≥48GB VRAM) to fully utilize context length and generation capacity. |
|
|
|
|
|
2. **Potential for Bias and Hallucinations** |
|
|
While tuned for alignment, some responses may still exhibit artifacts from pretraining biases or inaccuracies in edge cases. |
|
|
|
|
|
3. **Drift in Long Responses** |
|
|
Output may occasionally degrade in structure or accuracy across long generations. |
|
|
|
|
|
4. **Static Knowledge** |
|
|
Does not have real-time awareness or access to events or research developments post-training. |
|
|
|
|
|
5. **Creative Task Variability** |
|
|
While optimized for logic, its performance in narrative or subjective content may be inconsistent. |
