llm_advance_031_generation_config

Qwen3-4B AgentBench "023-Jinja-Heuristics" LoRA

This repository provides a highly optimized LoRA adapter fine-tuned from Qwen/Qwen3-4B-Instruct-2507. It is specifically engineered to achieve state-of-the-art performance on AgentBench (specifically ALFWorld and DBBench) by solving the catastrophic forgetting and format-collision problems inherent in multi-task agent fine-tuning.

This repository contains LoRA adapter weights only. The base model must be loaded separately.

🚀 Key Innovation: Jinja2 Contextual Routing & Heuristics Injection

The true power of this model lies not just in its weights, but in its custom tokenizer_config.json. We completely overrode the default chat_template using Jinja2 to act as an "Absolute Defense Shield" and a "Dynamic Heuristics Injector".

Depending on the user's prompt, the tokenizer automatically intercepts the input and injects task-specific System Prompts (Cheat Sheets) just before inference:

1. DB Bench (MySQL) Mode

When MySQL or SQL is detected in the prompt, the model is forced into a DB Agent persona with the following injected rules:

• Error Recovery: "If you encounter an SQL error (e.g., 'Unknown column'), DO NOT panic. Use Action: Operation to execute DESCRIBE table_name; and check the correct schema before retrying."
• Loop Prevention: "Never repeat the exact same invalid SQL."

2. ALFWorld (Household) Mode

When household or Interact with a is detected, the model is forced into an ALFWorld Agent persona:

• Format Override: Completely ignores the evaluation system's trap (THOUGHT:/ACTION:) and strictly enforces the stable Think:/Act: format.
• Exploration Logic: "If an action fails (Nothing happened), analyze why in your Think: step and choose a DIFFERENT action."
• Efficiency: "If you search a receptacle and do not find the target object, DO NOT search it again. Move to a different location."

📊 Training Configuration (The "Golden Ratio")

To maximize reasoning capabilities without exceeding the 4B model's capacity, we used a highly curated "Golden Ratio" dataset:

• Dataset: ALFWorld v5 Trajectories + DBBench Distilled (494 high-quality, noise-free trajectories).
• Method: LoRA (full precision base) via Unsloth.
• Loss Strategy: Loss is applied strictly to all assistant turns in the multi-turn trajectory, ignoring user/system prompts.

Hyperparameters:

• Max sequence length: 8192
• Epochs: 2
• Learning rate: 1e-6
• LoRA Rank (r): 64
• LoRA Alpha: 128
• Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj

💻 Usage

Because the magic is embedded in the Jinja2 chat_template, you must use this tokenizer to see the performance gains.

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base_model_id = "Qwen/Qwen3-4B-Instruct-2507"
adapter_id = "your_huggingface_id/your_model_name" # Change this to your actual repo ID

# 1. Load the customized tokenizer (CRITICAL)
tokenizer = AutoTokenizer.from_pretrained(adapter_id)

# 2. Load Base Model & Attach LoRA
model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(model, adapter_id)

# 3. Standard Inference (The Jinja2 template handles the routing automatically)
messages = [
    {"role": "user", "content": "You are a specialized MySQL database agent..."}
]
inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))