Update adapter: mix Magicoder code data (30% SQL / 70% code) to mitigate code-capability degradation

README.md

@@ -6,34 +6,85 @@ tags:
 - base_model:adapter:deepseek-ai/deepseek-coder-7b-instruct-v1.5
 - lora
 - sft
-- transformers
+- text-to-sql
 - trl
-licence: license
+license: apache-2.0
 pipeline_tag: text-generation
+datasets:
+- b-mc2/sql-create-context
+- ise-uiuc/Magicoder-OSS-Instruct-75K
 ---
 
 # Model Card for sql-adapter
 
-This model is a fine-tuned version of [deepseek-ai/deepseek-coder-7b-instruct-v1.5](https://huggingface.co/deepseek-ai/deepseek-coder-7b-instruct-v1.5).
-It has been trained using [TRL](https://github.com/huggingface/trl).
+LoRA adapter fine-tuned from [deepseek-ai/deepseek-coder-7b-instruct-v1.5](https://huggingface.co/deepseek-ai/deepseek-coder-7b-instruct-v1.5) for text-to-SQL generation, trained with [TRL](https://github.com/huggingface/trl) SFTTrainer.
+
+## What's new in this version
+
+The first version of this adapter was trained on 100% SQL data (`b-mc2/sql-create-context`) and showed catastrophic forgetting of general code generation ability (HumanEval pass@1 dropped from 60% to 50%, HumanEval+ from 50% to 40%).
+
+This version mixes in code-instruction data during training to mitigate that regression:
+
+| Dataset | Mix ratio | Purpose |
+|---|---|---|
+| [b-mc2/sql-create-context](https://huggingface.co/datasets/b-mc2/sql-create-context) | 30% | Text-to-SQL |
+| [ise-uiuc/Magicoder-OSS-Instruct-75K](https://huggingface.co/datasets/ise-uiuc/Magicoder-OSS-Instruct-75K) | 70% | Preserve general code generation ability |
+
+Mixing was done batch-wise via `datasets.interleave_datasets`. Learning rate was also lowered (2e-4 → 1e-4) and epochs increased (1 → 2) to compensate for the reduced SQL share.
+
+> Note: post-training benchmark numbers for this mixed-data run have not been re-measured yet. The SQL Exact Match / HumanEval numbers below are from the prior 100%-SQL run, kept here for reference; treat this release as a mitigation for known degradation rather than a verified improvement.
+
+## LoRA configuration
+
+```
+r: 16
+lora_alpha: 32
+lora_dropout: 0.05
+target_modules: [q_proj, k_proj, v_proj, o_proj]
+quantization: 4-bit NF4 (QLoRA), bf16 compute
+```
+
+## Prior version results (100% SQL training)
+
+| Metric | Base | SFT | Δ |
+|---|---|---|---|
+| SQL Exact Match (50 samples, sql-create-context) | 4% (2/50) | 78% (39/50) | +74% |
+| HumanEval pass@1 | 60% | 50% | -10% |
+| HumanEval+ (plus) | 50% | 40% | -10% |
 
 ## Quick start
 
 ```python
-from transformers import pipeline
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from peft import PeftModel
 
-question = "If you had a time machine, but could only go to the past or the future once and never return, which would you choose and why?"
-generator = pipeline("text-generation", model="None", device="cuda")
-output = generator([{"role": "user", "content": question}], max_new_tokens=128, return_full_text=False)[0]
-print(output["generated_text"])
-```
+base_model_id = "deepseek-ai/deepseek-coder-7b-instruct-v1.5"
+adapter_id = "albertkingdom/deepseek-coder-7b-instruct-sql-create-context-lora"
 
-## Training procedure
+model = AutoModelForCausalLM.from_pretrained(base_model_id, device_map="auto")
+tokenizer = AutoTokenizer.from_pretrained(base_model_id)
+model = PeftModel.from_pretrained(model, adapter_id)
 
- 
+messages = [{
+    "role": "user",
+    "content": """Given the database schema below, write a SQL query that answers the user's question.
+Only output the SQL query. Do not add any explanation.
 
+### Schema
+CREATE TABLE users (id INT, name VARCHAR(100), email VARCHAR(100))
 
-This model was trained with SFT.
+### Question
+Find all users with gmail addresses"""
+}]
+
+inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to(model.device)
+outputs = model.generate(inputs, max_new_tokens=200)
+print(tokenizer.decode(outputs[0], skip_special_tokens=True))
+```
+
+## Training procedure
+
+Trained with SFT (TRL `SFTTrainer`) on an interleaved SQL + code instruction dataset, 2 epochs, effective batch size 16, bf16, 4-bit QLoRA.
 
 ### Framework versions
 
@@ -46,10 +97,6 @@ This model was trained with SFT.
 
 ## Citations
 
-
-
-Cite TRL as:
-    
 ```bibtex
 @misc{vonwerra2022trl,
 	title        = {{TRL: Transformer Reinforcement Learning}},
@@ -59,4 +106,4 @@ Cite TRL as:
 	publisher    = {GitHub},
 	howpublished = {\url{https://github.com/huggingface/trl}}
 }
-```
+```