Upload 6 files

README.md

@@ -0,0 +1,11 @@
+# 🤖 DeepSeek-Coder LoRA Space
+Fine-tunes and serves `deepseek-ai/deepseek-coder-1.3b-base` using QLoRA.
+
+## Deploy on Hugging Face Spaces
+1. Create a new **Space** → Type: *Streamlit* → Hardware: *CPU Basic*
+2. Upload `deepseek-lora-space.zip`
+3. Add Secrets → `HF_TOKEN`, `WANDB_API_KEY`
+4. Space auto-builds; open the app!
+
+## Monthly auto-training
+GitHub Actions retrains monthly and rebuilds the Space if accuracy improves.

app.py

@@ -0,0 +1,13 @@
+import streamlit as st
+from inference import generate_response
+
+st.set_page_config(page_title="DeepSeek-Coder LoRA", page_icon="🤖")
+st.title("🤖 DeepSeek-Coder LoRA Playground")
+st.write("Fine-tuned on **NL2SH-ALFA** + user data using QLoRA")
+
+prompt = st.text_area("💬 Enter natural-language command:", height=150)
+if st.button("Generate Bash Command"):
+    with st.spinner("Thinking..."):
+        out = generate_response(prompt)
+        st.success(out)
+        st.code(out, language="bash")

inference.py

@@ -0,0 +1,14 @@
+from transformers import AutoTokenizer, AutoModelForCausalLM
+from peft import PeftModel
+
+BASE_MODEL = "deepseek-ai/deepseek-coder-1.3b-base"
+LORA_REPO  = "your-username/deepseek-lora-monthly"
+
+tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, trust_remote_code=True)
+base = AutoModelForCausalLM.from_pretrained(BASE_MODEL, trust_remote_code=True)
+model = PeftModel.from_pretrained(base, LORA_REPO)
+
+def generate_response(prompt:str)->str:
+    inputs = tokenizer(prompt, return_tensors="pt")
+    out = model.generate(**inputs, max_new_tokens=200)
+    return tokenizer.decode(out[0], skip_special_tokens=True)

monthly-train.yml

@@ -0,0 +1,34 @@
+name: Monthly Retrain + Deploy
+on:
+  schedule:
+    - cron: "0 3 1 * *"
+  workflow_dispatch:
+jobs:
+  retrain-deploy:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v4
+        with: { python-version: "3.10" }
+      - run: pip install -r requirements.txt
+      - env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          WANDB_API_KEY: ${{ secrets.WANDB_API_KEY }}
+        run: python train_lora.py
+      - name: Compare & deploy if improved
+        run: |
+          new=$(python -c "import json;print(1-json.load(open('out/metrics.json')).get('eval_loss',1))")
+          old=$(python -c "import json;print(json.load(open('best_score.json')).get('best_accuracy',0))" || echo 0)
+          echo "New acc=$new | Old=$old"
+          if (( $(echo "$new>$old"|bc -l) )); then
+            echo "{"best_accuracy":$new}" > best_score.json
+            git config user.name github-actions
+            git config user.email github-actions@github.com
+            git add best_score.json && git commit -m "update best score"
+            git push
+            huggingface-cli upload ./out/lora_adapters --repo your-username/deepseek-lora-monthly --token $HF_TOKEN
+            curl -X POST -H "Authorization: Bearer $HF_TOKEN"                  https://huggingface.co/api/spaces/your-username/deepseek-lora-space/rebuild
+            echo "🚀 Space rebuild triggered"
+          else
+            echo "No improvement. Skip deploy."
+          fi

requirements.txt

@@ -0,0 +1,11 @@
+transformers==4.46.0
+datasets
+accelerate
+wandb
+trl
+peft
+fastapi
+uvicorn
+streamlit
+huggingface_hub
+torch

train_lora.py

@@ -0,0 +1,50 @@
+import os, json, torch, wandb
+from transformers import (AutoModelForCausalLM, AutoTokenizer, Trainer,
+                          TrainingArguments, DataCollatorForLanguageModeling)
+from datasets import load_dataset
+from peft import LoraConfig, get_peft_model
+from huggingface_hub import HfApi
+
+HF_TOKEN = os.getenv("HF_TOKEN")
+WANDB_API_KEY = os.getenv("WANDB_API_KEY")
+wandb.login(key=WANDB_API_KEY)
+
+model_name = "deepseek-ai/deepseek-coder-1.3b-base"
+dataset = load_dataset("westenfelder/NL2SH-ALFA")
+
+tok = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
+def tok_fn(b): return tok([f"{n} => {bsh}" for n,bsh in zip(b['nl'],b['bash'])],
+                          truncation=True,padding="max_length",max_length=512)
+train, test = dataset["train"].map(tok_fn,batched=True), dataset["test"].map(tok_fn,batched=True)
+
+m = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16,
+                                         low_cpu_mem_usage=True, device_map="auto",
+                                         trust_remote_code=True)
+m.config.use_cache=False
+for p in m.parameters(): p.requires_grad=False
+
+cfg=LoraConfig(r=8,lora_alpha=16,target_modules=["q_proj","v_proj","k_proj","o_proj",
+                                                 "gate_proj","down_proj","up_proj"],
+               lora_dropout=0.05,bias="none",task_type="CAUSAL_LM")
+m=get_peft_model(m,cfg)
+coll=DataCollatorForLanguageModeling(tokenizer=tok,mlm=False)
+
+args=TrainingArguments(output_dir="./out",num_train_epochs=1,per_device_train_batch_size=1,
+                       gradient_accumulation_steps=8,learning_rate=2e-4,fp16=True,
+                       save_strategy="epoch",logging_steps=25,report_to=["wandb"])
+t=Trainer(model=m,args=args,train_dataset=train,eval_dataset=test,data_collator=coll)
+wandb.init(project="deepseek-qlora-monthly",name="deepseek-lite-run")
+t.train()
+
+metrics=t.evaluate(); acc=1-metrics.get("eval_loss",1)
+with open("out/metrics.json","w") as f: json.dump(metrics,f)
+wandb.log({"accuracy":acc})
+print(f"✅ Eval accuracy {acc:.4f}")
+
+ad="out/lora_adapters"; os.makedirs(ad,exist_ok=True)
+m.save_pretrained(ad); tok.save_pretrained(ad)
+artifact=wandb.Artifact("deepseek-lora-adapters","model"); artifact.add_dir(ad); wandb.log_artifact(artifact)
+
+api=HfApi(token=HF_TOKEN)
+api.upload_folder(folder_path=ad,repo_id="your-username/deepseek-lora-monthly",path_in_repo=".")
+print("✅ Uploaded to HF Hub")