Non-commercial access request
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This adapter is licensed under the Taimwe Non-Commercial License 1.0 (https://huggingface.co/Taimwe/qwen2.5-coder-7b-pro/blob/main/LICENSE). By requesting access you confirm that you will use it only for non-commercial purposes - personal projects, academic research, teaching or evaluation - and that your username and email address are shared with the model author. Commercial use requires a separate licence; see the LICENSE file.
Log in or Sign Up to review the conditions and access this model content.
Qwen2.5-Coder-7B-Pro (LoRA adapter)
A LoRA fine-tune of Qwen2.5-Coder-7B-Instruct, trained with
Unsloth + TRL on top of the 4-bit
quantised base unsloth/Qwen2.5-Coder-7B-Instruct-bnb-4bit.
This repository is the adapter only (~162 MB). If you want a ready-to-run model, use one of the exports:
| Export | Repository |
|---|---|
| Merged GGUF (Q4_K_M) | Taimwe/qwen2.5-coder-7b-pro-merged |
LoRA in GGUF (for llama.cpp --lora) |
Taimwe/qwen2.5-coder-7b-pro-F16-GGUF (private) |
Adapter configuration
| Setting | Value |
|---|---|
| PEFT type | LORA (plain, not rsLoRA) |
Rank r |
16 |
lora_alpha |
16 |
lora_dropout |
0.0 |
| Bias | none |
| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
| Base model | unsloth/Qwen2.5-Coder-7B-Instruct-bnb-4bit |
| Trained with | Unsloth / TRL, PEFT 0.20.0 |
Load the adapter
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base = AutoModelForCausalLM.from_pretrained(
"Qwen/Qwen2.5-Coder-7B-Instruct", torch_dtype="bfloat16", device_map="auto"
)
model = PeftModel.from_pretrained(base, "Taimwe/qwen2.5-coder-7b-pro")
tokenizer = AutoTokenizer.from_pretrained("Taimwe/qwen2.5-coder-7b-pro")
messages = [{"role": "user", "content": "Write a binary search in Rust."}]
inputs = tokenizer.apply_chat_template(
messages, add_generation_prompt=True, return_tensors="pt"
)
print(tokenizer.decode(model.generate(inputs, max_new_tokens=256)[0]))
This repository also ships chat_template.jinja, so apply_chat_template works
even though the base tokenizer config carries no template.
Merging into a full-precision base
The adapter was trained against a 4-bit base, but it must be merged into the bf16/fp16 base - PEFT cannot merge into a quantised model:
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
base = AutoModelForCausalLM.from_pretrained(
"Qwen/Qwen2.5-Coder-7B-Instruct", torch_dtype=torch.bfloat16, device_map="cpu"
)
merged = PeftModel.from_pretrained(base, "Taimwe/qwen2.5-coder-7b-pro").merge_and_unload()
merged.save_pretrained("qwen2.5-coder-7b-pro-merged-bf16")
AutoTokenizer.from_pretrained("Taimwe/qwen2.5-coder-7b-pro").save_pretrained(
"qwen2.5-coder-7b-pro-merged-bf16"
)
Publishing that bf16 merge as safetensors (rather than only GGUF) is what makes
this model servable by vLLM / TGI / SGLang and by hosted
Inference Providers - worth doing when you get a GPU.
Training and provenance
Reconstructed from the repository history and adapter_config.json:
| Item | Value |
|---|---|
| Training environment | Google Colab, Unsloth FastLanguageModel + TRL SFTTrainer |
| Upload path | push_to_hub/save_pretrained from that notebook - Hub commits "Upload model trained with Unsloth", 2026-09-20 06:35 UTC |
| Method | 4-bit QLoRA (base loaded in bnb-4bit, adapters trained in 16-bit) |
| Base | unsloth/Qwen2.5-Coder-7B-Instruct-bnb-4bit |
| Adapter | rank 16, alpha 16, dropout 0, all attention + MLP projections, PEFT 0.20.0 |
| Max sequence length | 32768 (as declared by the base tokenizer) |
| Licence | Apache-2.0 |
The GGUF exports made from this adapter are listed in the table at the top of this card; the merged Q4_K_M export was produced the same day (2026-09-20 09:20 UTC) and its Hub commits ("Trained with Unsloth", "- config", "- Ollama Modelfile") show it came from the same Unsloth notebook.
Known gaps
These values are not recorded in any of the repositories and cannot be recovered from the uploaded files - only the training notebook has them:
- Dataset(s) - name, revision and size. No dataset is published under this account, and no data is redistributed with the adapter.
- Number of steps / epochs
- Learning rate and schedule
- Max sequence length actually used during training
- GPU used (Colab free T4 vs Pro A100/L4) and total training time
- Evaluation results before/after fine-tuning
If the run was done with Unsloth in Colab they are all still in that notebook:
trainer.args (batch size, gradient accumulation, learning rate, epochs),
the max_seq_length passed to FastLanguageModel.from_pretrained, and the
dataset name given to SFTTrainer/train_on_responses_only. Publishing them is
what turns this from an upload into a reproducible fine-tune.
Limitations
Inherits every limitation of Qwen2.5-Coder-7B-Instruct, plus any drift introduced by the fine-tune. No benchmark comparison against the base model has been published, so the effect of the fine-tune is unverified.
Licence and attribution
Apache-2.0, following Qwen2.5-Coder-7B-Instruct (Apache-2.0). Trained 2x faster with Unsloth.
- Downloads last month
- 18