113 lines
4.2 KiB
Markdown
113 lines
4.2 KiB
Markdown
---
|
||
license: apache-2.0
|
||
datasets:
|
||
- TAUR-dev/STEPS__r1_4d_eval__mini_all
|
||
- TAUR-dev/STEPS__r1_8d_eval__v3_mini_all
|
||
- TAUR-dev/STEPS__r1_8d_eval__v4
|
||
- TAUR-dev/STEPS__r1_8d_eval__v3_4o
|
||
language:
|
||
- en
|
||
library_name: transformers
|
||
base_model:
|
||
- prithivMLmods/Qwen3-4B-ft-bf16
|
||
pipeline_tag: text-generation
|
||
tags:
|
||
- text-generation-inference
|
||
- trl
|
||
- moe
|
||
- code
|
||
- math
|
||
---
|
||
|
||
|
||
|
||
# Canum-Qwen3\_R1-4B-iCoT
|
||
|
||
> **Canum-Qwen3\_R1-4B-iCoT** is a precision-tuned variant of the Qwen3-4B architecture, explicitly aligned with **internal chain-of-thought (iCoT)** methodologies. Trained on the **TAUR-dev/STEPS\_\_r1\_4d\_eval\_\_mini\_all** dataset, this model excels in **long-form mathematical reasoning**, **progressive symbolic logic**, and **multi-stage problem decomposition**, all within a compact 4B parameter footprint.
|
||
|
||
> [!note]
|
||
GGUF : https://huggingface.co/prithivMLmods/Canum-Qwen3_R1-4B-iCoT-Q4_K_M-GGUF
|
||
|
||
|
||
## Key Features
|
||
|
||
1. **Internal Chain-of-Thought Reasoning (iCoT)**
|
||
Enables deeper logical progression through internally coherent reasoning steps, ideal for complex mathematical derivations and multivariable algebraic thinking.
|
||
|
||
2. **Dataset: TAUR-dev/STEPS\_\_r1\_4d\_eval\_\_mini\_all**
|
||
Fine-tuned using structured evaluation sequences to build resilience in multi-step problem solving and improve interpretability in math-focused tasks.
|
||
|
||
3. **Long Reasoning Paths in STEM Domains**
|
||
Suited for long-chain logical flows in geometry, number theory, calculus, and symbolic manipulation, including proofs and multi-stage equation solving.
|
||
|
||
4. **Lightweight Yet Capable (4B)**
|
||
Maintains strong reasoning and instruction-following abilities with lower computational cost compared to larger models, suitable for single-GPU deployments.
|
||
|
||
5. **Instruction-Following and Step-by-Step Alignment**
|
||
Follows complex instructions with multi-turn dependencies and provides granular output that aligns with internal steps used in the reasoning process.
|
||
|
||
6. **Technical Format Adaptability**
|
||
Outputs answers in clean Markdown, LaTeX, JSON, or table formats for academic, development, and notebook-based use cases.
|
||
|
||
## Quickstart with Transformers
|
||
|
||
```python
|
||
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
|
||
model_name = "prithivMLmods/Canum-Qwen3_R1-4B-iCoT"
|
||
|
||
model = AutoModelForCausalLM.from_pretrained(
|
||
model_name,
|
||
torch_dtype="auto",
|
||
device_map="auto"
|
||
)
|
||
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
||
|
||
prompt = "Use internal CoT to solve: A rectangle has a length that is 3 times its width. If the perimeter is 48 units, what are the dimensions?"
|
||
|
||
messages = [
|
||
{"role": "system", "content": "You are a reasoning assistant trained to use internal chain-of-thought (iCoT) for multi-step mathematical problems."},
|
||
{"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]
|
||
print(response)
|
||
```
|
||
|
||
## Intended Use
|
||
|
||
* Internal chain-of-thought (iCoT) problem solving
|
||
* Long-form symbolic math and algebraic derivations
|
||
* Curriculum-based step-by-step math tutoring
|
||
* Structured multi-turn reasoning in STEM domains
|
||
* Output generation in technical formats (LaTeX, Markdown)
|
||
|
||
## Limitations
|
||
|
||
* May require well-structured prompts for optimal reasoning output
|
||
* Smaller context length may limit extremely long multi-part problems
|
||
* Focused on precision reasoning, not creative or subjective writing
|
||
* Best used with prompt patterns that guide internal logical steps
|
||
|
||
## References
|
||
|
||
1. **TAUR-dev/STEPS\_\_r1\_4d\_eval\_\_mini\_all** – Dataset for structured math reasoning
|
||
2. **Internal CoT (iCoT)** – Progressive logical strategy for complex problems
|
||
3. [AIMO-2 Math Benchmark – OpenMathReasoning](https://arxiv.org/pdf/2504.16891)
|
||
4. [YaRN: Efficient Context Extension of LLMs](https://arxiv.org/pdf/2309.00071) |