File size: 4,862 Bytes
3e5299f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
---

language:
- code
- en
license: apache-2.0
datasets:
- bigcode/starcoderdata
library_name: transformers
pipeline_tag: text-generation
tags:
- code
- python
- code-completion
- fill-in-the-middle
- fim
- llama
---


# GrimmCoder

**GrimmCoder** is a ~107M-parameter LLaMA-style decoder-only transformer trained from scratch exclusively on Python source code for both **left-to-right completion** and **Fill-in-the-Middle (FIM)** code completion.

## Model Summary

| Property | Value |
|---|---|
| Architecture | LLaMA (`LlamaForCausalLM`, pre-norm, RoPE, SwiGLU, GQA, tied embeddings) |
| Parameters | ~106.98M total (~100.69M non-embedding) |
| Context Length | 2048 tokens |
| Hidden Size (`d_model`) | 768 |
| Layers | 16 |
| Attention Heads | 12 query heads / 4 KV heads (GQA) |
| FFN Intermediate Size | 2048 (SwiGLU) |
| Vocabulary | 8192 (byte-level BPE trained on Python) |
| Training Data | `bigcode/starcoderdata` (Python subset) |
| Tokens / Checkpoint | 7.93B tokens (step 61,500, validation loss 1.2321) |

## Special Tokens

| Token | ID | Purpose |
|---|---|---|
| `<|endoftext|>` | `0` | BOS / EOS / document delimiter |
| `<|fim_prefix|>` | `1` | FIM prefix marker |
| `<|fim_middle|>` | `2` | FIM middle marker (model generates from here) |
| `<|fim_suffix|>` | `3` | FIM suffix marker |
| `<|pad|>` | `4` | Padding token |

## Usage (`transformers`)

### 1. Left-to-Right Code Completion

```python

import torch

from transformers import AutoModelForCausalLM, AutoTokenizer



model_id = "Bayhakl1/GrimmCoder"

tokenizer = AutoTokenizer.from_pretrained(model_id)

model = AutoModelForCausalLM.from_pretrained(

    model_id,

    dtype=torch.bfloat16,

    device_map="auto",

)



prompt = """def binary_search(arr, target):

    lo, hi = 0, len(arr) - 1

"""

inputs = tokenizer(prompt, return_tensors="pt", add_special_tokens=False).to(model.device)

outputs = model.generate(

    **inputs,

    max_new_tokens=128,

    do_sample=True,

    temperature=0.2,

    top_p=0.95,

    eos_token_id=[0, 1, 2, 3, 4],

    pad_token_id=4,

)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

```

### 2. Fill-in-the-Middle (FIM)

The model was trained 50/50 on PSM and SPM Fill-in-the-Middle formatting. Note `add_special_tokens=False` when tokenizing raw prefix/suffix segments:

```python

prefix = "def add(a: int, b: int) -> int:\n    "

suffix = "\n\nprint(add(2, 3))\n"



# PSM format: <|fim_prefix|> prefix <|fim_suffix|> suffix <|fim_middle|>

fim_prompt = f"<|fim_prefix|>{prefix}<|fim_suffix|>{suffix}<|fim_middle|>"

inputs = tokenizer(fim_prompt, return_tensors="pt", add_special_tokens=False).to(model.device)



outputs = model.generate(

    **inputs,

    max_new_tokens=64,

    do_sample=True,

    temperature=0.2,

    top_p=0.95,

    eos_token_id=[0, 1, 2, 3, 4],

    pad_token_id=4,

)

middle_ids = outputs[0][inputs["input_ids"].shape[1]:]

middle = tokenizer.decode(middle_ids, skip_special_tokens=True)

print(prefix + middle + suffix)

```

## Evaluation (HumanEval)

Evaluated on [`openai/openai_humaneval`](https://huggingface.co/datasets/openai/openai_humaneval) (164 Python problems) using `transformers` (`bfloat16`):

| Prompt Mode | Decoding | `pass@1` | Solved |
|---|---|---|---|
| Left-to-Right (`prompt.rstrip("\n")`, ending on `"""`) | greedy (`do_sample=False`, `max_new_tokens=128`) | **5.49%** | 9 / 164 |
| Left-to-Right (`prompt.rstrip("\n")`, ending on `"""`) | `temperature=0.2`, `top_k=50`, `top_p=0.95`, `repetition_penalty=1.1` | **4.88%** | 8 / 164 |
| Fill-in-the-Middle (PSM: `<\|fim_prefix\|>{prompt}<\|fim_suffix\|>\n<\|fim_middle\|>`) | greedy (`do_sample=False`, `max_new_tokens=384`) | **4.27%** | 7 / 164 |
| Left-to-Right (raw `prompt` ending on `"""\n`) | greedy (`do_sample=False`, `max_new_tokens=384`) | **0.00%** | 0 / 164 |

> **Notes:**
> - As a base completion model pretrained on raw `starcoderdata`, prompts ending directly on `"""\n` often resemble empty stub/interface files (predicting `<|endoftext|>` or starting another `def`). Stripping the trailing newline so the prompt ends on `"""` (or prompting in FIM-PSM mode) cues the model to emit the indented function body instead.
> - In `bfloat16` (`transformers`), minor intermediate precision rounding can affect borderline tokens (for example, `HumanEval/48` passes in `float32` while `bfloat16` scores `9/164` / `5.49%` greedy). Also keep `repetition_penalty=1.0` when using `transformers` `generate()`, as HF applies `repetition_penalty` across the prompt docstring as well.

## Limitations

- Trained exclusively on Python source code; not intended as a general natural-language chat or instruction-following model.
- Base / completion model: responds best to code prompts and FIM markers rather than conversational Q&A.