en license: mit library_name: transformers tags:
reasoning
axolotl
r1 base_model:
meta-llama/Llama-3.2-3B-Instruct datasets:
ServiceNow-AI/R1-Distill-SFT pipeline_tag: text-generation model-index:
name: DeepSeek-R1-Distill-Llama-3B results:
- task: type: text-generation name: Text Generation dataset: name: IFEval (0-Shot) type: HuggingFaceH4/ifeval args: num_few_shot: 0 metrics:
  - type: inst_level_strict_acc and prompt_level_strict_acc value: 70.93 name: strict accuracy source: url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=suayptalha/DeepSeek-R1-Distill-Llama-3B name: Open LLM Leaderboard
- task: type: text-generation name: Text Generation dataset: name: BBH (3-Shot) type: BBH args: num_few_shot: 3 metrics:
  - type: acc_norm value: 21.45 name: normalized accuracy source: url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=suayptalha/DeepSeek-R1-Distill-Llama-3B name: Open LLM Leaderboard
- task: type: text-generation name: Text Generation dataset: name: MATH Lvl 5 (4-Shot) type: hendrycks/competition_math args: num_few_shot: 4 metrics:
  - type: exact_match value: 20.92 name: exact match source: url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=suayptalha/DeepSeek-R1-Distill-Llama-3B name: Open LLM Leaderboard
- task: type: text-generation name: Text Generation dataset: name: GPQA (0-shot) type: Idavidrein/gpqa args: num_few_shot: 0 metrics:
  - type: acc_norm value: 1.45 name: acc_norm source: url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=suayptalha/DeepSeek-R1-Distill-Llama-3B name: Open LLM Leaderboard
- task: type: text-generation name: Text Generation dataset: name: MuSR (0-shot) type: TAUR-Lab/MuSR args: num_few_shot: 0 metrics:
  - type: acc_norm value: 2.91 name: acc_norm source: url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=suayptalha/DeepSeek-R1-Distill-Llama-3B name: Open LLM Leaderboard
- task: type: text-generation name: Text Generation dataset: name: MMLU-PRO (5-shot) type: TIGER-Lab/MMLU-Pro config: main split: test args: num_few_shot: 5 metrics:
  - type: acc value: 21.98 name: accuracy source: url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=suayptalha/DeepSeek-R1-Distill-Llama-3B name: Open LLM Leaderboard

DeepSeek-R1-Distill-Llama-3B

This model is the distilled version of DeepSeek-R1 on Llama-3.2-3B with R1-Distill-SFT dataset.

See axolotl config

yaml

base_model: unsloth/Llama-3.2-3B-Instruct
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: true
load_in_4bit: false
strict: false

chat_template: llama3
datasets:
  - path: ./custom_dataset.json
    type: chat_template
    conversation: chatml
    ds_type: json

add_bos_token: true
add_eos_token: true
use_default_system_prompt: false

special_tokens:
  bos_token: "<|begin_of_text|>"
  eos_token: "<|eot_id|>"
  pad_token: "<|eot_id|>"
  additional_special_tokens:
    - "<|begin_of_text|>"
    - "<|eot_id|>"

adapter: lora
lora_model_dir:
lora_r: 16
lora_alpha: 32
lora_dropout: 0.1
lora_target_linear: true

hub_model_id: suayptalha/DeepSeek-R1-Distill-Llama-3B

sequence_len: 2048
sample_packing: false
pad_to_sequence_len: true
micro_batch_size: 2
gradient_accumulation_steps: 8
num_epochs: 1
learning_rate: 2e-5
optimizer: paged_adamw_8bit
lr_scheduler: cosine

train_on_inputs: false
group_by_length: false
bf16: false
fp16: true
tf32: false

gradient_checkpointing: true
flash_attention: false

logging_steps: 50
warmup_steps: 100
saves_per_epoch: 1

output_dir: ./finetune-sft-results
save_safetensors: true

Prompt Template

You can use Llama3 prompt template while using the model:

Llama3

<|start_header_id|>system<|end_header_id|>
{system}<|eot_id|>

<|start_header_id|>user<|end_header_id|>
{user}<|eot_id|>

<|start_header_id|>assistant<|end_header_id|>
{assistant}<|eot_id|>

Example usage:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "suayptalha/DeepSeek-R1-Distill-Llama-3B",
    device_map="auto"
)

tokenizer = AutoTokenizer.from_pretrained("suayptalha/DeepSeek-R1-Distill-Llama-3B")

SYSTEM_PROMPT = """Respond in the following format:
<think>
You should reason between these tags.
</think>

Answer goes here...

Always use <think> </think> tags even if they are not necessary.
"""

messages = [
    {"role": "system", "content": SYSTEM_PROMPT},
    {"role": "user", "content": "Which one is larger? 9.11 or 9.9?"},
]
inputs = tokenizer.apply_chat_template(
    messages,
    tokenize = True,
    add_generation_prompt = True,
    return_tensors = "pt",
).to("cuda")
output = model.generate(input_ids=inputs, max_new_tokens=256, use_cache=True, temperature=0.7)
decoded_output = tokenizer.decode(output[0], skip_special_tokens=False)
print(decoded_output)

Output:

<think>
First, I need to compare the two numbers 9.11 and 9.9. 

Next, I'll analyze each number. The first digit after the decimal point in 9.11 is 1, and in 9.9, it's 9. 

Since 9 is greater than 1, 9.9 is larger than 9.11.
</think>

To determine which number is larger, let's compare the two numbers:

**9.11** and **9.9**

1. **Identify the Decimal Places:**
   - Both numbers have two decimal places.
   
2. **Compare the Tens Place (Right of the Decimal Point):**
   - **9.11:** The tens place is 1.
   - **9.9:** The tens place is 9.
   
3. **Conclusion:**
   - Since 9 is greater than 1, the number with the larger tens place is 9.9.
   
**Answer:** **9.9** is larger than **9.11**.

Suggested system prompt:

Respond in the following format:
<think>
You should reason between these tags.
</think>

Answer goes here...

Always use <think> </think> tags even if they are not necessary.

Parameters

lr: 2e-5
epochs: 1
batch_size: 16
optimizer: paged_adamw_8bit

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	23.27
IFEval (0-Shot)	70.93
BBH (3-Shot)	21.45
MATH Lvl 5 (4-Shot)	20.92
GPQA (0-shot)	1.45
MuSR (0-shot)	2.91
MMLU-PRO (5-shot)	21.98

Support

RichardErkhov/suayptalha_-_DeepSeek-R1-Distill-Llama-3B-4bits

Original model description: