Hatim2221
/

Fikr-7B-Reasoning

Text Generation

chain-of-thought

Model card Files Files and versions

Hatim2221 commited on Aug 29

Commit

2dbf6fe

·

verified ·

1 Parent(s): 1c591ec

Update README.md

Files changed (1) hide show

README.md +2 -1

README.md CHANGED Viewed

@@ -31,7 +31,8 @@ Both models were evaluated on the full test split (**1,319 questions**) of the *
 | Model | Architecture | Arabic Reasoning Format | Arabic-GSM8K Accuracy | Delta (Δ) |
 | :--- | :--- | :--- | :---: | :---: |
 | `Qwen/Qwen2.5-7B-Instruct` (Base) | 7B Dense | Standard CoT | **72.40%**[cite: 3] | Baseline |
-| **`Fikr-7B-Reasoning` (Ours)** | **7B Dense** | **`<think>` Native CoT** | **70.81%**[cite: 4] | **-1.59%** |
 ### Key Takeaways
 * **Accuracy Trade-off:** While the model strictly enforces the `<think>` tag reasoning structure, it experienced a minor **-1.59%** regression in raw exact-match accuracy compared to the strong Qwen 2.5 7B baseline[cite: 3, 4]. This is a common phenomenon in first-pass SFT reasoning models and typically requires a follow-up Reinforcement Learning (RL) phase (like DPO or PPO) to align the structured thoughts with correct final answers.

 | Model | Architecture | Arabic Reasoning Format | Arabic-GSM8K Accuracy | Delta (Δ) |
 | :--- | :--- | :--- | :---: | :---: |
 | `Qwen/Qwen2.5-7B-Instruct` (Base) | 7B Dense | Standard CoT | **72.40%**[cite: 3] | Baseline |
+| **`Fikr-7B-Reasoning` (Ours)** | **7B Dense** | **`<think>` Native CoT** | **71.95%**[cite: 4] | **-45%** |
 ### Key Takeaways
 * **Accuracy Trade-off:** While the model strictly enforces the `<think>` tag reasoning structure, it experienced a minor **-1.59%** regression in raw exact-match accuracy compared to the strong Qwen 2.5 7B baseline[cite: 3, 4]. This is a common phenomenon in first-pass SFT reasoning models and typically requires a follow-up Reinforcement Learning (RL) phase (like DPO or PPO) to align the structured thoughts with correct final answers.