mayank-dubey-ai/enterprise-100-qwen3-8b-steering-vectors

Enterprise-100 Qwen3-8B Steering Vectors

Reusable activation-steering artifacts from the Enterprise-100 database experiment. They steer Qwen/Qwen3-8B between direct SQL and Python sqlite3/pandas solutions without changing model weights.

This is an activation artifact repository, not a fine-tuned checkpoint. Load the base Qwen3-8B model separately and apply a vector through a forward hook.

Core result

The principal direction is:

python_minus_sql = mean(h_python - h_sql)

At generation time:

h' = h + alpha × v

• alpha < 0: biases toward direct SQL
• alpha = 0: leaves the base model unchanged
• alpha > 0: biases toward Python database code

The most stable demonstration uses layer 16. The vector has 4,096 float32 components, matching the Qwen3-8B residual-stream width.

Contents

Artifact	Purpose
vectors/python_vs_sql_engineer_layer_{8,12,16,20,24,28}.pt	Layer sweep of the primary Python-minus-SQL direction
vectors/full_code_modality_layer_{12,16,20}.pt	Broader fenced-code modality direction
vectors/format_python.pt	Python-versus-SQL formatting direction
vectors/persona_executive.pt	Executive-versus-engineer persona direction
diagnostics/python_vs_sql_diagnostics.json	Norms and within-pair cosine diagnostics
config.json	Machine-readable compatibility and intervention metadata

Vector quality diagnostics

Four contrast pairs were used for the principal direction. Mean cosine alignment of each pair-specific difference with the mean vector remained high across the layer sweep:

Layer	Vector norm	Mean pair cosine
8	30.92	0.908
12	41.62	0.897
16	49.91	0.900
20	72.46	0.912
24	147.56	0.908
28	258.40	0.893

High alignment indicates a consistent contrast direction across the four examples; it does not by itself prove causal usefulness. Causal behavior is tested by intervening during generation and executing the resulting programs.

Minimal usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = "Qwen/Qwen3-8B"
artifact = torch.load(
    "vectors/python_vs_sql_engineer_layer_16.pt",
    map_location="cpu",
    weights_only=True,
)

tokenizer = AutoTokenizer.from_pretrained(base_model)
model = AutoModelForCausalLM.from_pretrained(
    base_model,
    dtype=torch.bfloat16,
    device_map="auto",
).eval()

layer = model.model.layers[artifact["layer"]]
vector = artifact["vector"].to(model.device, model.dtype)
alpha = 0.75  # positive -> Python; negative -> SQL

def steering_hook(_module, _inputs, output):
    hidden = output[0] if isinstance(output, tuple) else output
    steered = hidden.clone()
    steered[:, -1, :] += alpha * vector
    return (steered, *output[1:]) if isinstance(output, tuple) else steered

handle = layer.register_forward_hook(steering_hook)
try:
    # Call model.generate(...) here.
    pass
finally:
    handle.remove()

Use the complete scripts and 100-table SQLite fixture in the linked dataset repository for reproducible prompts, schema retrieval, execution, and scoring.

Benchmark summary

The published 10-query benchmark covers joins across 2–6 tables. In the fast execution run:

• Negative/SQL generation executed successfully on 10/10 queries.
• Positive/Python generation executed successfully on 8/10 queries.
• SQL and Python results sometimes differed in row count or semantics even when both programs executed.

Execution success is therefore not semantic correctness. The repository keeps raw outputs and row counts visible so these differences can be audited.

Limitations

• Extracted from only four Python/SQL contrast pairs.
• Validated on one model family and checkpoint: Qwen/Qwen3-8B.
• Direction magnitude changes substantially with layer; alpha values are not directly comparable across layers.
• Forward hooks are implementation-specific and can break across model architecture changes.
• Steering changes probability distributions; it does not guarantee valid or safe code.
• Never execute generated database or Python code against production systems without sandboxing, validation, and authorization.

Related resources

• Dataset and full reproduction code
• Interactive benchmark explorer
• Base model