---
license: bigscience-openrail-m
library_name: transformers
tags:
- code
- gpt_bigcode
datasets:
- nuprl/MultiPL-T
metrics:
- code_eval
model-index:
- name: MultiPLCoder-15b-OCaml
  results:
  - task:
      type: text-generation
    dataset:
      name: MultiPL-HumanEval (Lua)
      type: nuprl/MultiPL-E
    metrics:
    - type: pass@1
      value: 0.31
      name: pass@1
      verified: true
    - type: pass@1
      value: 0.21
      name: pass@1
      verified: true
    - type: pass@1
      value: 0.199
      name: pass@1
      verified: true
---
# MultiPLCoder-15b

15 billion parameter version of MultiPLCoder, a set of StarCoder-based models finetuned on the [MultiPL-T dataset](https://huggingface.co/datasets/nuprl/MultiPL-T).
These models are state-of-the-art at low-resource languages, such as: Lua, Racket, and OCaml.

This 15 billion parameter model is the most capable of the MultiPLCoder family. However, it requires a dedicated GPU for inference.
For a smaller model that fits on the CPU, check out [MultiPLCoder-1b](https://huggingface.co/nuprl/MultiPLCoder-1b).


## Language Revision Index

This is the revision index for the best-performing models for their respective langauge.

| Langauge      | Revision ID | Epoch |
| ------------- | ----------- | ----- |
| Lua           | `6069aa54dd554404dd18fccdf5dedd56b8088e74`         | 4    |
| Racket        | `f0c77c06482f436f469007f20d731cb9dd73d609`         | 8    |
| OCaml         | `e7babda985786810707200ff885df6105de7dc56`         | 4    |

## Usage

To utilize one of the models in this repository, you must first select a commit revision for that model from the table above.
For example, to use the Lua model:
```py
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("nuprl/MultiPLCoder-15b")
lua_revision="6069aa54dd554404dd18fccdf5dedd56b8088e74"
model = AutoModelForCausalLM.from_pretrained("nuprl/MultiPLCoder-15b", revision=lua_revision).cuda()
```

Note that the model's default configuration does not enable caching, therefore you must specify to use the cache on generation.
```py
toks = tokenizer.encode("-- Fibonacci iterative", return_tensors="pt").cuda()
out = model.generate(toks, use_cache=True,  do_sample=True, temperature=0.2, top_p=0.95, max_length=256)
print(tokenizer.decode(out[0], skip_special_tokens=True))
```
```
-- Fibonacci iterative.
local function fib_iterative(n)
    if n == 0 or n == 1 then
        return n
    end
    local previous, current = 0, 1
    for _ = 2, n do
        previous, current = current, current + previous
    end
    return current
end
```