Model Card for mypo-qwen2.5-coder-1.5b-sft

SFT LoRA adapter for Qwen/Qwen2.5-Coder-1.5B-Instruct, trained on the chosen (type-hinted) responses of joshuasundance/mypo-4k-rfc. The goal is to teach the base model to emit modern, fully-annotated Python code.

This is a PEFT/LoRA adapter, not a merged model. Load the base model first, then attach the adapter.

TL;DR — does it work?

Two complementary measurements are reported — both published, both reproducible.

Batched eval (n=150 stratified held-out validation prompts, run 2026-04-22-qwen2.5-1.5b):

metric	base	this adapter (sft)	gold (chosen)
parse rate	97.3 %	100 %	100 %
ruff pass rate	93.3 %	96.0 %	100 %
black pass rate	12.0 %	97.3 %	98.0 %
mypy --strict pass rate	6.0 %	92.7 %	100 %
mean annotation slot coverage	0.000	0.953	0.955
fully-annotated function fraction	0.000	0.893	0.898
mean ruff violations / sample	0.47	0.07	0.00
mean mypy errors / sample	2.30	0.13	0.00
preference win rate vs base	—	100 %	—
preference win rate vs rejected	—	100 %	—
preference win rate vs chosen	—	49.0 % (tie)	—

Single-prompt validation (n=30, batch=1, no padding — realistic one-user inference):

metric	base	this adapter (sft)
parse rate	93.3 %	100 %
ruff pass rate	90.0 %	93.3 %
black pass rate	6.7 %	100 %
mypy --strict pass rate	0.0 %	73.3 %
mean annotation slot coverage	0.000	0.971

The core claim — base never annotates, SFT almost always does — survives under single-prompt decoding. The batched and single-prompt validations are retained as two different in-domain measurement regimes, but we no longer attribute their gap specifically to left-padding or batching as a general causal explanation. Full characterization is in joshuasundance/mypo-training; single-prompt artifacts are in single-prompt-validation/single-prompt-2026-04-23T002137Z/.

External benchmark note: on the latest canonical full HumanEval+ run (n=164), this adapter reaches 97 / 164 = 59.1 % pass@1 on base tests and 86 / 164 = 52.4 % on plus tests, which is below the untuned Qwen base (112 / 164, 99 / 164). So SFT improves the narrow type-hinting objective, but not broad HumanEval+ correctness.

Quick start

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

base_id = "Qwen/Qwen2.5-Coder-1.5B-Instruct"
adapter_id = "joshuasundance/mypo-qwen2.5-coder-1.5b-sft"

tokenizer = AutoTokenizer.from_pretrained(base_id)
base = AutoModelForCausalLM.from_pretrained(base_id, torch_dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(base, adapter_id)

messages = [{"role": "user", "content": "Write a typed Python function that returns the nth Fibonacci number."}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=256, do_sample=False)
print(tokenizer.decode(out[0, inputs.shape[-1]:], skip_special_tokens=True))

Training

Trained with TRL SFTTrainer on a single NVIDIA A10G via Hugging Face Jobs. Training script: mypo_sft_train.py. Job id: 69e8a4522aa1660eaffa894d.

Data

• Dataset: joshuasundance/mypo-4k-rfc (train + validation concatenated → 6361 examples).
• Format: DPO (prompt, chosen, rejected) collapsed to SFT conversational, keeping only chosen.

Hyperparameters

Group	Setting
Quantization	4-bit NF4 + double-quant, bnb_4bit_compute_dtype=bfloat16
LoRA	r=256, alpha=16, dropout=0.05, target_modules="all-linear", task_type=CAUSAL_LM
Optimization	paged_adamw_8bit, lr=2e-4, cosine schedule, warmup_steps=100
Batching	per_device_train_batch_size=1, gradient_accumulation_steps=8 (effective 8)
Schedule	3 epochs, max_length=2048, packing=off
Precision	bf16, gradient checkpointing on
Seed	42

Final training metrics (from job logs)

Metric	Value
train_runtime	8339.5 s (~2h 19m)
train_loss	0.3679
mean_token_accuracy (final step)	0.9038
grad_norm (final step)	~0.66

Environmental impact

CodeCarbon telemetry from the training run (emissions.csv in this repo):

Metric	Value
Duration	8339.5 s
Energy consumed	0.4721 kWh
CO₂e emissions	0.1743 kg
GPU energy / power	0.3079 kWh / 132.9 W avg
CPU energy / power	0.0434 kWh / 19.4 W avg
RAM energy / power	0.1209 kWh / 54.0 W
Hardware	1 × NVIDIA A10G, AMD EPYC 7R32 (48 vCPU), 187 GB RAM
Region	AWS us-east-1 (Virginia), PUE 1.0
Tracker	codecarbon 3.2.6, tracking_mode=machine

Framework versions

• TRL 1.2.0, Transformers 5.5.4, PyTorch 2.11.0, Datasets 4.8.4, Tokenizers 0.22.2, PEFT ≥ 0.12, bitsandbytes ≥ 0.44.

License

Apache 2.0 (inherits from the Qwen2.5-Coder-1.5B-Instruct base model).

Citations

CodeCarbon (emissions tracking)

Emissions for this run were measured with CodeCarbon v3.2.6.

@software{codecarbon,
  author  = {Benoit Courty and Victor Schmidt and Sasha Luccioni and Goyal-Kamal and MarionCoutarel and Boris Feld and Jérémy Lecourt and LiamConnell and Amine Saboni and Inimaz and supatomic and Mathilde Léval and Luis Blanche and Alexis Cruveiller and Ouminasara and Franklin Zhao and Aditya Joshi and Alexis Bogroff and Hugues de Lavoreille and Niko Laskaris and Edoardo Abati and Douglas Blank and Ziyao Wang and Armin Catovic and Marc Alencon and Michał Stęchły and Christian Bauer and Lucas Otávio N. de Araújo and JPW and MinervaBooks},
  title   = {{CodeCarbon: Estimate and track carbon emissions from machine learning computing}},
  year    = 2024,
  doi     = {10.5281/zenodo.11171501},
  url     = {https://github.com/mlco2/codecarbon}
}

TRL

@software{vonwerra2020trl,
  title   = {{TRL: Transformer Reinforcement Learning}},
  author  = {von Werra, Leandro and Belkada, Younes and Tunstall, Lewis and Beeching, Edward and Thrush, Tristan and Lambert, Nathan and Huang, Shengyi and Rasul, Kashif and Gallouédec, Quentin},
  license = {Apache-2.0},
  url     = {https://github.com/huggingface/trl},
  year    = 2020
}