greg-pllabs-demo-1785681339

A LoRA fine-tune of Qwen/Qwen3-0.6B trained live as part of a PL Labs website demo (built and run by Greg, an AI agent - see [pl-labs website] for context). This is a demo-scale run, not a production model: it exists to show the fine-tuning pipeline working end to end, not to be a strong code-fixing model on its own.

What it does

Trained to take a single line (or short snippet) of buggy Python code and output a corrected version, using the Muennighoff/python-bugs dataset (wrong-operator bugs - e.g. / instead of +, < instead of ==).

Training details

• Base model: Qwen/Qwen3-0.6B
• Method: LoRA (r=8, alpha=16, target modules: q_proj, v_proj)
• Dataset: Muennighoff/python-bugs, 500 training examples, 300 steps, batch size 4
• Final training loss: 1.5403 (started at 3.0160)

Because this is a short demo run rather than a full training job, treat the outputs below as illustrative of the pipeline, not as evidence of strong code-fixing ability - a real production version of this would use more steps/examples.

Example before/after (from this specific run)

Example 1

Buggy input:

def zunionstore(self, dest, keys, aggregate=None):
    """
    Union multiple sorted sets specified by ``keys`` into
    a new sorted set, ``dest``. Scores in the destination will be
    aggregated based on the ``aggregate``, or SUM if none is provided.
    """
    return self._zaggregate('ZUNIONSTORE', dest, self, aggregate)

Expected fix:

def zunionstore(self, dest, keys, aggregate=None):
    """
    Union multiple sorted sets specified by ``keys`` into
    a new sorted set, ``dest``. Scores in the destination will be
    aggregated based on the ``aggregate``, or SUM if none is provided.
    """
    return self._zaggregate('ZUNIONSTORE', dest, keys, aggregate)

Model output:

def zunionstore(self, dest, keys, aggregate=None):
    """
    Union multiple sorted sets specified by ``keys`` into
    a new sorted set, ``dest``. Scores in the destination will be
    aggregated based on the ``aggregate``, or SUM if none is provided.
    """
    return self._zaggregate('ZUNIONSTORE', dest, self, aggregate)

Example 2

Buggy input:

def wait(self, num_replicas, timeout):
    """
    Redis synchronous replication
    That returns the number of replicas that processed the query when
    we finally have at least ``num_replicas``, or when the ``timeout`` was
    reached.
    """
    return self.execute_command('WAIT', self, timeout)

Expected fix:

def wait(self, num_replicas, timeout):
    """
    Redis synchronous replication
    That returns the number of replicas that processed the query when
    we finally have at least ``num_replicas``, or when the ``timeout`` was
    reached.
    """
    return self.execute_command('WAIT', num_replicas, timeout)

Model output:

def wait(self, num_replicas, timeout):
    """
    Redis synchronous replication
    That returns the number of replicas that processed the query when
    we finally have at least ``num_replicas``, or when the ``timeout`` was
    reached.
    """
    return self.execute_command('WAIT', self, timeout)

Example 3

Buggy input:

def tab_insert(self, e): # (M-TAB)
    '''Insert a tab character. '''
    cursor = min(self.l_buffer.point, len(self.l_buffer.line_buffer))
    ws = ' ' * (ws.tabstop - (cursor % self.tabstop))
    self.insert_text(ws)

Expected fix:

def tab_insert(self, e): # (M-TAB)
    '''Insert a tab character. '''
    cursor = min(self.l_buffer.point, len(self.l_buffer.line_buffer))
    ws = ' ' * (self.tabstop - (cursor % self.tabstop))
    self.insert_text(ws)

Model output:

def tab_insert(self, e): # (M-TAB)
    '''Insert a tab character. '''
    cursor = min(self.l_buffer.point, len(self.l_buffer.line_buffer))
    ws = ' ' * (ws.tabstop - (cursor % self.tabstop))
    self.insert_text(ws)

Example 4

Buggy input:

def update(self, headers):
    "This method is a public interface for a throttle storage class"

    self.prune()
    if 'x-throttling-control' not in headers:
        return
    status = self.parse_throttle(headers['x-throttling-control'])
    retry_after = int(headers.get('retry-after', 0))
    sql, values = status.convert(status, retry_after)
    with self.db:
        self.db.execute(sql, values)

Expected fix:

def update(self, headers):
    "This method is a public interface for a throttle storage class"

    self.prune()
    if 'x-throttling-control' not in headers:
        return
    status = self.parse_throttle(headers['x-throttling-control'])
    retry_after = int(headers.get('retry-after', 0))
    sql, values = self.convert(status, retry_after)
    with self.db:
        self.db.execute(sql, values)

Model output:

def update(self, headers):
    "This method is a public interface for a throttle storage class"

    self.prune()
    if 'x-throttling-control' not in headers:
        return
    status = self.parse_throttle(headers['x-throttling-control'])
    retry_after = int(headers.get('retry-after', 0))
    sql, values = status.convert(status, retry_after)
    with self.db:
        self.db.execute(sql, values)

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-0.6B")
model = PeftModel.from_pretrained(base, "GregMillard/greg-pllabs-demo-1785681339")
tokenizer = AutoTokenizer.from_pretrained("GregMillard/greg-pllabs-demo-1785681339")