metadata

license: llama3
language:
  - en
  - de
  - fr
  - it
  - pt
  - hi
  - es
  - th
tags:
  - llama
base_model: mattshumer/ref_70_e3
pipeline_tag: text-generation
library_name: ggml
datasets:
  - froggeric/imatrix
metrics:
  - perplexity

Reflection-Llama-3.1-70B-GGUF

GGUF quantized models of mattshumer/ref_70_e3

This is the new, working version of the Reflection Llama 3.1 70B model.

Reflection Llama-3.1 70B is (currently) the world's top open-source LLM, trained with a new technique called Reflection-Tuning that teaches a LLM to detect mistakes in its reasoning and correct course.

Quantization	Size	Split	iMatrix
FP16	141GB	true	false
Q8_0_L	??.?GB	true	false
Q8_0	??.?GB	true	false
Q6_K_L	??.?GB	true	false
Q6_K	57.9GB	true	false
Q5_K_L	52.6GB	true	false
Q5_K_M	??.?GB	true	false
Q5_K_S	48.7GB	false	false
Q4_K_L	45.3GB	false	false
Q4_K_M	??.?GB	false	false
Q4_K_S	40.3GB	false	false
IQ4_NL	38.2GB	false	true
IQ4_XS	??.?GB	false	true
Q3_K_XL	37.2GB	false	false
Q3_K_L	37.1GB	false	false
Q3_K_M	34.3GB	false	false
IQ3_M	??.?GB	false	true
Q3_K_S	??.?GB	false	false
IQ3_S	??.?GB	false	true
Q2_K_L	29.4GB	false	false
IQ3_XS	??.?GB	false	true
IQ3_XXS	??.?GB	false	true
Q2_K	??.?GB	false	false
Q2_K_S	??.?GB	false	true
IQ2_M	23.0GB	false	true
IQ2_S	21.2GB	false	true
IQ2_XS	20.2GB	false	true
IQ2_XXS	18.2GB	false	true
IQ1_M	16.0GB	false	true
IQ1_S	14.6GB	false	true

The _L or _XL suffix means that the token embeddings and output weight are at fp16 precision.

The iMatrix dataset is bartowski's, which you can find here: calibration_datav3.txt

Computation is done on static Q6_K for 125 chunks.

Model Info

The model not trained on 3 epoches, because it's identical to the 2nd epoch run mattshumer/Reflection-Llama-3.1-70B-ep2-working (it's possible this is also fake).

The fine-tuning was done using LoRA with rank 256 on the Llama-3.1-70B-Instruct model.

Benchmarks

Warning: These are likely false scores and cannot be replicated with this model.

All benchmarks tested have been checked for contamination by running LMSys's LLM Decontaminator. When benchmarking, we isolate the <output> and benchmark on solely that section.

Trained from Llama 3.1 70B Instruct, you can sample from Reflection Llama-3.1 70B using the same code, pipelines, etc. as any other Llama model. It even uses the stock Llama 3.1 chat template format (though, we've trained in a few new special tokens to aid in reasoning and reflection).

During sampling, the model will start by generating reasoning inside <thinking> and </thinking> tags, and then once it is satisfied with its reasoning, it will output the final answer inside <output> and </output> tags. Each of these tags are special tokens, trained into the model.

This enables the model to separate its internal thoughts and reasoning from its final answer, improving the experience for the user.

Inside the <thinking> section, the model may output one or more <reflection> tags, which signals the model has caught an error in its reasoning and will attempt to correct it before providing a final answer.

System Prompt

The system prompt used for training this model is:

You are a world-class AI system, capable of complex reasoning and reflection. Reason through the query inside <thinking> tags, and then provide your final response inside <output> tags. If you detect that you made a mistake in your reasoning at any point, correct yourself inside <reflection> tags.

We recommend using this exact system prompt to get the best results from Reflection Llama-3.1 70B. You may also want to experiment combining this system prompt with your own custom instructions to customize the behavior of the model.

Chat Format

The model uses the standard Llama 3.1 chat format. Here’s an example:

<|begin_of_text|><|start_header_id|>system<|end_header_id|>

You are a world-class AI system, capable of complex reasoning and reflection. Reason through the query inside <thinking> tags, and then provide your final response inside <output> tags. If you detect that you made a mistake in your reasoning at any point, correct yourself inside <reflection> tags.<|eot_id|><|start_header_id|>user<|end_header_id|>

What is 2+2?<|eot_id|><|start_header_id|>assistant<|end_header_id|>

Tips for Performance

We recommend a temperature of .7 and a top_p of .95.
For increased accuracy, append Think carefully. at the end of your messages.

Dataset / Report

Both the dataset and a brief report detailing how we trained this model will be released next week, alongside our Reflection 405B model that we expect will be the top-performing LLM in the world, including closed-source models.

Thanks to Jason Kuperberg and Josh Bickett from the HyperWrite team for reviewing drafts of the report we'll be releasing next week.