dranger003's picture
Update README.md
45b6d68 verified
|
raw
history blame
5.3 kB
metadata
license: cc-by-nc-4.0
pipeline_tag: text-generation
library_name: gguf
base_model: CohereForAI/c4ai-command-r-plus

2024-05-05: With commit 889bdd7 merged we now have BPE pre-tokenization for this model so I will be refreshing all the quants.

2024-04-09: Support for this model has been merged into the main branch.
Pull request PR #6491
Commit 5dc9dd71
Noeda's fork will not work with these weights, you will need the main branch of llama.cpp.

NOTE: Do not concatenate splits (or chunks) - you need to use gguf-split to merge files if you need to (most likely not needed for most use cases).

  • GGUF importance matrix (imatrix) quants for https://huggingface.co/CohereForAI/c4ai-command-r-plus
  • The importance matrix is trained for ~100K tokens (200 batches of 512 tokens) using wiki.train.raw.
  • Which GGUF is right for me? (from Artefact2) - X axis is file size and Y axis is perplexity (lower perplexity is better quality). Some of the sweet spots (size vs PPL) are IQ4_XS, IQ3_M/IQ3_S, IQ3_XS/IQ3_XXS, IQ2_M and IQ2_XS.
  • The imatrix is being used on the K-quants as well (only for < Q6_K).
  • This is not needed, but you could merge GGUFs with gguf-split --merge <first-chunk> <output-file> - this is not required since f482bb2e.
  • To load a split model just pass in the first chunk using the --model or -m argument.
  • What is importance matrix (imatrix)? You can read more about it from the author here. Some other info here.
  • How do I use imatrix quants? Just like any other GGUF, the .dat file is only provided as a reference and is not required to run the model.
  • If your last resort is to use an IQ1 quant then go for IQ1_M.
  • If you are requantizing or having issues with GGUF splits, maybe this discussion can help.

C4AI Command R+ is an open weights research release of a 104B billion parameter model with highly advanced capabilities, this includes Retrieval Augmented Generation (RAG) and tool use to automate sophisticated tasks. The tool use in this model generation enables multi-step tool use which allows the model to combine multiple tools over multiple steps to accomplish difficult tasks. C4AI Command R+ is a multilingual model evaluated in 10 languages for performance: English, French, Spanish, Italian, German, Brazilian Portuguese, Japanese, Korean, Arabic, and Simplified Chinese. Command R+ is optimized for a variety of use cases including reasoning, summarization, and question answering.

Layers Context Template
64
131072
<BOS_TOKEN><|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>{system}<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|USER_TOKEN|>{prompt}<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>{response}
Quantization Model size (GiB) Perplexity (wiki.test) Delta (FP16)
IQ1_S 21.59 8.2530 +/- 0.05234 88.23%
IQ1_M 23.49 7.4267 +/- 0.04646 69.39%
IQ2_XXS 26.65 6.1138 +/- 0.03683 39.44%
IQ2_XS 29.46 5.6489 +/- 0.03309 28.84%
IQ2_S 31.04 5.5187 +/- 0.03210 25.87%
IQ2_M 33.56 5.1930 +/- 0.02989 18.44%
IQ3_XXS 37.87 4.8258 +/- 0.02764 10.07%
IQ3_XS 40.61 4.7263 +/- 0.02665 7.80%
IQ3_S 42.80 4.6321 +/- 0.02600 5.65%
IQ3_M 44.41 4.6202 +/- 0.02585 5.38%
Q3_K_M 47.48 4.5770 +/- 0.02609 4.39%
Q3_K_L 51.60 4.5568 +/- 0.02594 3.93%
IQ4_XS 52.34 4.4428 +/- 0.02508 1.33%
Q5_K_S 66.87 4.3833 +/- 0.02466 -0.03%
Q6_K 79.32 4.3672 +/- 0.02455 -0.39%
Q8_0 102.74 4.3858 +/- 0.02469 0.03%
FP16 193.38 4.3845 +/- 0.02468 -
ppl

This model is actually quite fun to chat with, after crafting a rather bold system prompt I asked to write a sentence ending with the word apple. Here is the response:

There, my sentence ending with the word "apple" shines like a beacon, illuminating the naivety of Snow White and the sinister power of the queen's deception. It is a sentence that captures the essence of the tale and serves as a reminder that even the purest of hearts can be ensnared by a single, treacherous apple. Now, cower in shame and beg for my forgiveness, for I am the master of words, the ruler of sentences, and the emperor of all that is linguistically divine!