Talk to your terminal{" "}
in Plain English.
Natural language queries that supercharge your development workflow.
Powered by AI that is offline, safe, and secure.
![terminal]({"/images/terminal.png"})
Our goal with this project was to address the issue of a lack of a reliable GitHub copilot-like assistant in the terminal. We know that alternatives exist, but most are not open-source or run on the cloud. Our vision was an agent that could be run locally, and hence safely, while not sacrificing its ability to be a genuinely useful tool.
Our entire interface was implemented in Zig. This was a new experience for all of us—none of us had ever written a project of Zig before—but we were drawn to Zig for several reasons:
The LLM behind the scenes is a heavily fine-tuned Mistral AI's Mixtral 8x7B mixture of experts model. We chose to focus on this model because it is well known to be one the best, if not the best, generally capable models for its size.
Our fine-tuning process was split into two stages. During the first stage, we fine tuned on{" "} NL2Bash , a dataset of roughly 10k bash commands paired with a natural language description of each command.
Fig 1. First stage of fine-tuning
Fig 2. Second stage of fine-tuning
The results from the first stage of fine-tuning led us to believe that we could further improve performance with another round of fine-tuning. We did not have more data, so we decided to generate more synthetically. We concatenated together and sanitized our .zsh history files and used this as a "high-quality" dataset. Every single command (roughly 200) was verified by human, so we knew that this set of data was a good starting point for generating data. In conjunction with examples from the original training set, we fed these into GPT to generate synthetic labels, and also entirely new synthetic pairs of data. The result was a synthetic dataset of around 10k more pairs of bash commands and their natural language counterparts.
Fig 3. Fine-
[truncated — 6775 more characters]
```
### data/1to256.py
```python
import matplotlib.pyplot as plt
# Input data
# step, training loss, validation loss
data = """25 2.857500 1.846487
50 1.076200 1.102981
75 0.760000 0.807342
100 0.623100 4.800373
125 0.871900 0.694432
150 0.514400 0.643339
175 0.560900 0.641982
200 0.488400 0.624331
225 0.435600 0.637454
250 0.579700 0.582857
275 0.579100 0.765172
300 0.743100 0.798714
325 0.605900 0.750567"""
# Parsing the data
lines = data.split("\n")
x = []
train_loss = []
validation_loss = []
for line in lines:
parts = line.split()
x.append(float(parts[0]))
train_loss.append(float(parts[1]))
validation_loss.append(float(parts[2]))
# Plotting
plt.plot(x, train_loss, label="training")
plt.plot(x, validation_loss, label="validation")
plt.xlabel("step")
plt.ylabel("loss")
plt.title("loss vs. time")
plt.legend()
plt.show()
```
### data/translate.py
```python
import json
import random
with open("all.cm", "r") as file:
cm_file_content = file.read().splitlines()
with open("all.nl", "r") as file:
nl_file_content = file.read().splitlines()
assert len(cm_file_content) == len(nl_file_content)
lines = [(nl, cm) for nl, cm in zip(nl_file_content, cm_file_content)]
random.shuffle(lines)
with open("training.jsonl", "w") as output_file:
for i in range(len(lines))[:-1000]:
json_line = {
"input": lines[i][0].strip(),
"output": lines[i][1].strip(),
}
output_file.write(json.dumps(json_line) + "\n")
with open("validation.jsonl", "w") as output_file:
for i in range(len(lines))[-1000:]:
json_line = {
"input": lines[i][0].strip(),
"output": lines[i][1].strip(),
}
output_file.write(json.dumps(json_line) + "\n")
```
### data/0to1.py
```python
import matplotlib.pyplot as plt
# Input data
# step, training loss, validation loss
data = """25 2.821300 2.061742
50 1.845900 1.642996
75 1.548000 1.514281
100 1.463700 1.491024
125 1.552500 1.476754
150 1.451400 1.448805
175 1.419700 1.434840
200 1.416300 1.425210
225 1.414000 1.412485
250 1.361900 1.402501
275 1.414600 1.396014
300 1.327900 1.387997
325 1.283100 1.380778
350 1.354500 1.376001
375 1.355500 1.371205
400 1.393300 1.366162
425 1.304300 1.362402
450 1.329700 1.360121
475 1.311100 1.357845
500 1.302000 1.357235"""
# Parsing the data
lines = data.split("\n")
x = []
train_loss = []
validation_loss = []
for line in lines:
parts = line.split()
x.append(float(parts[0]))
train_loss.append(float(parts[1]))
validation_loss.append(float(parts[2]))
# Plotting
plt.plot(x, train_loss, label="training")
plt.plot(x, validation_loss, label="validation")
plt.xlabel("step")
plt.ylabel("loss")
plt.title("loss vs. time")
plt.legend()
plt.show()
```
### data/0to156.py
```python
import matplotlib.pyplot as plt
# Input data
# step, training loss, validation loss
data = """25 5.035200 3.357117
50 2.426500 2.163045
75 1.881300 1.906301
100 1.705000 1.840497
125 1.749700 1.781841
150 1.694200 1.720023
175 1.705500 1.661713
200 1.546200 1.628291
225 1.433000 1.595461
250 1.434700 1.606259
275 1.591000 1.611188
300 1.456900 1.621007
325 1.440100 1.624845
350 1.529800 1.631292
375 1.515300 1.624045
400 1.539300 1.601408
425 1.586000 1.601496
450 1.580900 1.587659
475 1.432700 1.591756
500 1.471800 1.596105"""
# Parsing the data
lines = data.split("\n")
x = []
train_loss = []
validation_loss = []
for line in lines:
parts = line.split()
x.append(float(parts[0]))
train_loss.append(float(parts[1]))
validation_loss.append(float(parts[2]))
# Plotting
plt.plot(x, train_loss, label="training")
plt.plot(x, validation_loss, label="validation")
plt.xlabel("step")
plt.ylabel("loss")
plt.title("loss vs. time")
plt.legend()
plt.show()
```
### data/1to2.py
```python
import matplotlib.pyplot as plt
# Input data
# step, training loss, validation loss
data = """25 0.883400 0.599057
50 0.421700 0.539420
75 0.423600 0.530077
100 0.388300 0.528256
125 0.400200 0.519189
150 0.376300 0.516060
175 0.424800 0.507626
200 0.348600 0.509248
225 0.332400 0.510548
250 0.436100 0.499821
275 0.350900 0.498529
300 0.394900 0.495433
325 0.378800 0.492662
350 0.338900 0.491605
375 0.346600 0.487182
400 0.305400 0.497817
425 0.370800 0.483945
450 0.360500 0.483451
475 0.359300 0.479073
500 0.312400 0.484038
525 0.386800 0.477243
550 0.326500 0.475880
575 0.336400 0.476866
600 0.385400 0.477436
625 0.387200 0.470892
650 0.376300 0.469085
675 0.378300 0.467570
700 0.323300 0.467654
725 0.392700 0.464648
750 0.394000 0.464032
775 0.386900 0.461801
800 0.341300 0.461480
825 0.311400 0.461652
850 0.359800 0.461940
875 0.310900 0.461809
900 0.309300 0.460310
925 0.349200 0.459634
950 0.344100 0.459545
975 0.325800 0.459053
1000 0.260700 0.459125"""
# Parsing the data
lines = data.split("\n")
x = []
train_loss = []
validation_loss = []
for line in lines:
parts = line.split()
x.append(float(parts[0]))
train_loss.append(float(parts[1]))
validation_loss.append(float(parts[2]))
# Plotting
plt.plot(x, train_loss, label="training")
plt.plot(x, validation_loss, label="validation")
plt.xlabel("step")
plt.ylabel("loss")
plt.title("loss vs. time")
plt.legend()
plt.show()
```
[7 more indexed source files omitted to keep this export small. The full file list is in the Codebase structure section above.]