Model Card: Ministral-3-GLM-Distilled-R1
Model Overview
Model Name: Ministral-3-GLM-Distilled-R1
Base Model: mistralai/Ministral-3B
Model Type: Causal Language Model (Text-Only)
Architecture: Ministral-3 (3B parameters) – vision layers frozen
Training Method: Supervised Fine-Tuning (SFT) + Group Relative Policy Optimization (GRPO)
Model Description
This model is a distilled reasoning engine fine-tuned from GLM-5.2 logic puzzles and reinforced with GRPO to produce concise, structured reasoning. It uses a [THINK]...[/THINK] block for reasoning and a [SOLUTION]...[/SOLUTION] block for the final answer.
The model is designed for tasks requiring:
- Logical reasoning
- Mathematical problem-solving
- Step-by-step explanation
- Concise, structured outputs
Model Details
| Attribute | Value |
|---|---|
| Architecture | Ministral-3 (3B) with LoRA adapters |
| Context Length | 4096 tokens |
| Vocabulary | 131,072 tokens |
| Training Data | GLM-5.2 Logic Puzzles (860 examples) + 100 synthetic reasoning examples |
| Total Training Examples | 919 (after filtering) |
| LoRA Rank | 8 |
| Trainable Parameters | 24.7M (0.64% of total) |
| Training Steps (SFT) | 120 (2 epochs) |
| Training Steps (GRPO) | 50 |
Training Procedure
Phase 1: Supervised Fine-Tuning (SFT)
- Dataset:
ianncity/GLM-5.2-Logic-Puzzles+ synthetic[THINK]examples - Filtering: Removed examples >2048 tokens
- Loss: Converged from ~11.2 to ~0.38
- Format:
[THINK]...[/THINK]+ final answer
Phase 2: GRPO Reinforcement Learning
- Dataset:
open-r1/Big-Math-RL-Verified-Processed(100 examples) - Reward Functions:
- Format compliance (1.0 for exact
[THINK]/[/THINK]tags) - Correctness (string match – improved in v2)
- Format compliance (1.0 for exact
- Hyperparameters:
- Learning rate: 5e-6
num_generations: 4- KL beta: 0.001
- Temperature: 0.8 (training)
Usage
Default System Prompt (Critical)
# HOW YOU SHOULD THINK AND ANSWER
First draft your thinking process (inner monologue) until you arrive at a response. Do NOT overthink, immediately stop if you are checking your answer more than 2 times. Format your response using Markdown, and use LaTeX for any mathematical equations. Write both your thoughts and the response in the same language as the input.
Your thinking process must follow the template below:
[THINK]Your concise thoughts, like working through an exercise on scratch paper. Be as casual and as long as you want until you are confident to generate the response to the user.[/THINK]
Here, provide a self-contained response.
Put your final short answer inside [SOLUTION] and [/SOLUTION]
Recommended Sampling Parameters
| Parameter | Value |
|---|---|
temperature | 0.5 |
top_k | 40 |
top_p | 0.9 |
min_p | 0.1 |
repetition_penalty | 1.2 |
max_new_tokens | 4096 |
Strengths and Limitations
Strengths
- Concise Reasoning: Outputs are ~1500 tokens – 2× shorter than the SFT version.
- Structured Format: Consistently produces
[THINK]and[SOLUTION]blocks. - Logic Puzzle Mastery: Handles ambiguity and multiple solutions well.
- Self‑Correction: Can catch and fix errors in reasoning.
Limitations
- Math Handling: Limited – reward function was string‑based (string match), not math‑aware.
- Sampling Sensitivity: Output quality depends on specific sampling parameters.
- Looping: Can over‑verify at low temperature; system prompt is critical.
- Hallucination: May mis‑evaluate statements (e.g., calling
trueasfalse).
When to Use This Model
| Use Case | Recommended |
|---|---|
| Logic puzzles | ✅ Yes |
| Step‑by‑step reasoning | ✅ Yes |
| Concise explanations | ✅ Yes |
| Mathematical computation | ⚠️ Limited (use with math-verify post‑processing) |
| Long‑form reasoning | ✅ Yes (with 4096 token limit) |
Training Costs
| Phase | Steps | Duration (estimated) |
|---|---|---|
| SFT | 120 | ~2 hours |
| GRPO | 50 | ~1.5 hours |
Considerations
-
This model is designed for reasoning tasks and should not be used for:
- Medical diagnosis
- Financial advice
- Legal decisions
- Generating harmful or deceptive content
-
Hallucinations are possible. Always verify critical outputs manually.
Uploaded finetuned model
- Developed by: ertghiu256
- License: apache-2.0
- Finetuned from model : ertghiu256/ministral-3-3b-glm-distillation-test
This mistral3 model was trained 2x faster with Unsloth and Huggingface's TRL library.