mohdnihalll03/ qwen2.5-coder-1.5b-codeslm-nihal

56 Downloads Updated 6 days ago

A QLoRA fine-tune of Qwen2.5-Coder-1.5B into a terse, code-first Python assistant — 78% shorter answers, trained on a single 8 GB GPU.

ollama run mohdnihalll03/qwen2.5-coder-1.5b-codeslm-nihal

curl http://localhost:11434/api/chat \
  -d '{
    "model": "mohdnihalll03/qwen2.5-coder-1.5b-codeslm-nihal",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='mohdnihalll03/qwen2.5-coder-1.5b-codeslm-nihal',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'mohdnihalll03/qwen2.5-coder-1.5b-codeslm-nihal',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Models

Name

1 model

Size / Usage

Context

Input

qwen2.5-coder-1.5b-codeslm-nihal:latest

986MB · 32K context window · Text · 6 days ago

qwen2.5-coder-1.5b-codeslm-nihal:latest

986MB

32K

Text

Readme

qwen2.5-coder-1.5b-CodeSLM-Nihal

A QLoRA fine-tune of Qwen2.5-Coder-1.5B-Instruct into a terse, code-first Python assistant — trained on a single 8 GB consumer GPU (RTX 5060) as a phase-by-phase learning project.

The base model is a capable coder but chronically verbose (every answer trails a prose essay). This fine-tune trains it to reply with correct, minimal code and nothing else.

Headline result — 78% shorter answers

On 18 held-out prompts (greedy decoding), average output length dropped from 272 to 60 tokens (−78%) while core algorithm correctness was preserved.

Training curve — clean convergence, no overfitting

Checkpoint (epoch)	Validation loss
0.13	0.5157
0.38	0.5020
0.88	0.4927
1.00	0.4928

Final train loss 0.4906; validation mean token-accuracy ~85.4%.

Training details


Base	Qwen/Qwen2.5-Coder-1.5B-Instruct
Method	QLoRA (4-bit NF4 + LoRA), completion-only loss
Dataset	sahil2801/CodeAlpaca-20k → Qwen ChatML (19,020 / 1,002)
LoRA	r=16, α=32, dropout=0.05, all 7 proj modules
Trainable	18.46M params (~1.18%)
Schedule	1 epoch, 1,189 steps, LR 2e-4 cosine + 3% warmup
Hardware	RTX 5060 (8 GB), ~50 min, peak VRAM 3.32 GB

Usage

ollama run mohdnihalll03/qwen2.5-coder-1.5b-codeslm-nihal

Recommended: temperature 0.2. Full weights, adapter, and GGUFs on Hugging Face: https://huggingface.co/MohdNihal03/qwen2.5-coder-1.5b-CodeSLM-Nihal

Limitations

Fine-tuning changed style far more than correctness. A few subtle base-model bugs persist and one answer regressed — review generated code before use. Python-focused; 1.5B + 4-bit is not a substitute for a large frontier model.

# qwen2.5-coder-1.5b-CodeSLM-Nihal

A **QLoRA fine-tune of Qwen2.5-Coder-1.5B-Instruct** into a **terse, code-first Python assistant** —
trained on a single 8 GB consumer GPU (RTX 5060) as a phase-by-phase learning project.

The base model is a capable coder but chronically verbose (every answer trails a prose essay). This
fine-tune trains it to reply with **correct, minimal code and nothing else.**

## Headline result — 78% shorter answers

![Output length: 78% shorter after fine-tuning](https://huggingface.co/MohdNihal03/qwen2.5-coder-1.5b-CodeSLM-Nihal/resolve/main/token_reduction.png)

On 18 held-out prompts (greedy decoding), average output length dropped **from 272 to 60 tokens
(−78%)** while core algorithm correctness was preserved.

## Training curve — clean convergence, no overfitting

![Training vs. validation loss](https://huggingface.co/MohdNihal03/qwen2.5-coder-1.5b-CodeSLM-Nihal/resolve/main/loss_curve.png)

| Checkpoint (epoch) | Validation loss |
|---|---|
| 0.13 | 0.5157 |
| 0.38 | 0.5020 |
| 0.88 | 0.4927 |
| 1.00 | 0.4928 |

Final train loss **0.4906**; validation mean token-accuracy **~85.4%**.

## Training details

| | |
|---|---|
| Base | Qwen/Qwen2.5-Coder-1.5B-Instruct |
| Method | QLoRA (4-bit NF4 + LoRA), completion-only loss |
| Dataset | sahil2801/CodeAlpaca-20k → Qwen ChatML (19,020 / 1,002) |
| LoRA | r=16, α=32, dropout=0.05, all 7 proj modules |
| Trainable | 18.46M params (~1.18%) |
| Schedule | 1 epoch, 1,189 steps, LR 2e-4 cosine + 3% warmup |
| Hardware | RTX 5060 (8 GB), ~50 min, peak VRAM 3.32 GB |

## Usage

```
ollama run mohdnihalll03/qwen2.5-coder-1.5b-codeslm-nihal
```

Recommended: temperature 0.2. Full weights, adapter, and GGUFs on Hugging Face:
**https://huggingface.co/MohdNihal03/qwen2.5-coder-1.5b-CodeSLM-Nihal**

## Limitations

Fine-tuning changed *style* far more than *correctness*. A few subtle base-model bugs persist and one
answer regressed — **review generated code before use.** Python-focused; 1.5B + 4-bit is not a
substitute for a large frontier model.

Paste, drop or click to upload images (.png, .jpeg, .jpg, .svg, .gif)