107 lines
3.8 KiB
Markdown
107 lines
3.8 KiB
Markdown
|
---
|
|||
|
|
license: apache-2.0
|
|||
|
|
base_model: Qwen/Qwen2.5-Coder-1.5B-Instruct
|
|||
|
|
tags:
|
|||
|
|
- code
|
|||
|
|
- python
|
|||
|
|
- qlora
|
|||
|
|
- lora
|
|||
|
|
- peft
|
|||
|
|
- qwen2
|
|||
|
|
- text-generation
|
|||
|
|
datasets:
|
|||
|
|
- sahil2801/CodeAlpaca-20k
|
|||
|
|
language:
|
|||
|
|
- en
|
|||
|
|
pipeline_tag: text-generation
|
|||
|
|
library_name: transformers
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
# 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**,
|
|||
|
|
built as a hands-on, phase-by-phase fine-tuning learning project on a single 8 GB consumer GPU
|
|||
|
|
(NVIDIA RTX 5060).
|
|||
|
|
|
|||
|
|
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
|
|||
|
|
|
|||
|
|
|
|||
|
|
|
|||
|
|
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
|
|||
|
|
|
|||
|
|
|
|||
|
|
|
|||
|
|
| Checkpoint (epoch) | Validation loss |
|
|||
|
|
|---|---|
|
|||
|
|
| 0.13 | 0.5157 |
|
|||
|
|
| 0.38 | 0.5020 |
|
|||
|
|
| 0.88 | 0.4927 |
|
|||
|
|
| 1.00 | 0.4928 |
|
|||
|
|
|
|||
|
|
Train and validation loss tracked each other the entire run — clean convergence, **no overfitting.**
|
|||
|
|
Final train loss **0.4906**; validation mean token-accuracy **~85.4%**.
|
|||
|
|
|
|||
|
|
> These are training/next-token metrics plus a qualitative 18-prompt comparison. No standardized code
|
|||
|
|
> benchmark (HumanEval/MBPP) was run.
|
|||
|
|
|
|||
|
|
## Training details
|
|||
|
|
|
|||
|
|
| | |
|
|||
|
|
|---|---|
|
|||
|
|
| Base | `Qwen/Qwen2.5-Coder-1.5B-Instruct` |
|
|||
|
|
| Method | QLoRA (4-bit NF4 base + LoRA), completion-only loss |
|
|||
|
|
| Dataset | `sahil2801/CodeAlpaca-20k` → Qwen ChatML (19,020 train / 1,002 val) |
|
|||
|
|
| LoRA | r=16, α=32, dropout=0.05, targets q/k/v/o/gate/up/down |
|
|||
|
|
| Trainable params | 18.46M (~1.18%) |
|
|||
|
|
| Schedule | 1 epoch, 1,189 steps, LR 2e-4 cosine + 3% warmup |
|
|||
|
|
| Batch | 2 × 8 grad-accum = effective 16 |
|
|||
|
|
| Optimizer | paged_adamw_8bit, bf16, gradient checkpointing, max_len 1024 |
|
|||
|
|
| Hardware | 1× RTX 5060 (8 GB), ~50 min, peak VRAM 3.32 GB |
|
|||
|
|
|
|||
|
|
## Files in this repo
|
|||
|
|
|
|||
|
|
- **Root** — merged fp16 model (Transformers format); load directly with `from_pretrained`.
|
|||
|
|
- **`adapter/`** — the standalone LoRA adapter (~36 MB) to apply onto the base yourself.
|
|||
|
|
- **`gguf/`** — `…-f16.gguf` (full precision) and `…-Q4_K_M.gguf` (~986 MB) for llama.cpp / Ollama.
|
|||
|
|
|
|||
|
|
## Usage
|
|||
|
|
|
|||
|
|
### Transformers
|
|||
|
|
```python
|
|||
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|||
|
|
m = "MohdNihal03/qwen2.5-coder-1.5b-CodeSLM-Nihal"
|
|||
|
|
tok = AutoTokenizer.from_pretrained(m)
|
|||
|
|
model = AutoModelForCausalLM.from_pretrained(m, device_map="auto")
|
|||
|
|
msgs = [{"role": "user", "content": "Write a Python function that reverses a string without slicing."}]
|
|||
|
|
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
|
|||
|
|
print(tok.decode(model.generate(ids, max_new_tokens=128)[0][ids.shape[1]:], skip_special_tokens=True))
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
### Ollama
|
|||
|
|
```
|
|||
|
|
ollama run mohdnihalll03/qwen2.5-coder-1.5b-codeslm-nihal
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
Prompt format: Qwen2.5 ChatML. Suggested: `temperature 0.2`, stop on `<|im_start|>` / `<|im_end|>`.
|
|||
|
|
|
|||
|
|
## Limitations
|
|||
|
|
|
|||
|
|
- **Style over correctness:** fine-tuning changed formatting far more than correctness. A few subtle
|
|||
|
|
base-model bugs persist (an email-regex character-class quirk; a `@timer` decorator missing
|
|||
|
|
`functools.wraps`), and one bracket-matching answer regressed to a logic bug on empty-stack input.
|
|||
|
|
**Review generated code before use.**
|
|||
|
|
- Python-focused; 1.5B params + 4-bit quantization — not a substitute for a large frontier model.
|
|||
|
|
- Occasional instruction drift (`print` vs `return`, tabulation vs memoization).
|
|||
|
|
|
|||
|
|
## Attribution
|
|||
|
|
|
|||
|
|
- Base: Qwen2.5-Coder-1.5B-Instruct (© Alibaba Cloud, Apache-2.0)
|
|||
|
|
- Data: `sahil2801/CodeAlpaca-20k`
|
|||
|
|
- Fine-tuned by **Nihal** as a QLoRA learning project (TRL / PEFT / bitsandbytes).
|