Model Card for mypo-qwen2.5-coder-1.5b-dpo-v2

⚠️ Superseded by mypo-qwen2.5-coder-1.5b-dpo-v3. v3 fixes the v2 no-op issue (matched LoRA α=r=256, lr=5e-5, β=0.3, warm-started from SFT, published as a merged model). This v2 adapter is preserved as a reproducibility artifact and case study in how DPO training telemetry can look healthy while generations remain unchanged.

DPO LoRA adapter for Qwen/Qwen2.5-Coder-1.5B-Instruct, trained on joshuasundance/mypo-4k-rfc, where chosen = type-hinted Python and rejected = unhinted Python. v2 uses a higher LoRA rank and all-linear targeting compared to v1.

This is a PEFT/LoRA adapter, not a merged model.

TL;DR — honest assessment

Training telemetry looked great (rewards/accuracies → 1.0, rewards/margins → ~0.44), but post-hoc characterization shows this adapter barely shifts generation behavior from the base model under greedy decoding. Two measurements confirm the no-op under different decoding regimes:

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

metric	base	this adapter (dpo-v2)	SFT adapter	gold (chosen)
parse rate	97.3 %	97.3 %	100 %	100 %
ruff pass rate	93.3 %	94.0 %	96.0 %	100 %
black pass rate	12.0 %	12.0 %	97.3 %	98.0 %
mypy --strict pass rate	6.0 %	6.0 %	92.7 %	100 %
mean annotation slot coverage	0.000	0.000	0.953	0.955
mean ruff violations / sample	0.47	0.46	0.07	0.00
mean mypy errors / sample	2.30	2.35	0.13	0.00
preference win rate vs chosen	—	0 %	49.0 %	—
preference win rate vs base	—	50 % (tie)	100 %	—

Single-prompt validation (n=30, batch=1, no padding):

metric	base	this adapter (dpo-v2)	SFT adapter
mypy --strict pass rate	0.0 %	0.0 %	73.3 %
mean annotation slot coverage	0.000	0.000	0.971
black pass rate	6.7 %	6.7 %	100 %

v2 is indistinguishable from base under both decoding regimes — batched and single-prompt — confirming the no-op is not an artifact of batching, tokenization, or adapter-loading. The adapter is active (outputs differ from base in non-annotation ways), but the policy shift in annotation dimensions is effectively zero.

Why the training metrics look strong but generation doesn't change

DPO optimizes a ranking objective on log p(chosen) − log p(rejected). Achieving rewards/accuracies = 1.0 only requires the log-ratio to tilt in the right direction — it can be satisfied by infinitesimal weight movements when the LoRA scaling factor is small.

Here the effective LoRA scale is alpha / r = 16 / 256 = 0.0625, and the DPO learning rate is 1e-6 (200× lower than the SFT recipe's 2e-4). The product — tiny deltas × tiny scale × small lr — left the adapter weights too small to change argmax generations under greedy decoding, even though they were enough to move relative log-probabilities.

For a dpo-v3, either raise lora_alpha to match r (→ alpha=256, scale=1.0) or switch to rslora=True, and increase lr to 1e-5–5e-5. Those are documented remedies in the "LoRA without regret" literature.

Full characterization is published in joshuasundance/mypo-training.

External benchmark note: on the latest canonical full HumanEval+ run (n=164), this adapter reaches 110 / 164 = 67.1 % pass@1 on base tests and 97 / 164 = 59.1 % on plus tests, versus the untuned Qwen base at 112 / 164 and 99 / 164. So v2 remains close to base out of domain, but it still does not beat the base model on general 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-dpo-v2"

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 DPOTrainer on a single NVIDIA A10G via Hugging Face Jobs. Training script: mypo_dpo_train_v2.py. Job id: 69e8a5b0d2fd2eb837d766b1.

Data

• Dataset: joshuasundance/mypo-4k-rfc (train + validation concatenated → 6361 preference pairs).
• Format: native DPO (prompt, chosen, rejected).

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=1e-6, cosine schedule, warmup_steps=100
Batching	per_device_train_batch_size=1, gradient_accumulation_steps=8 (effective 8)
DPO	beta=0.1, loss_type="sigmoid", reference adapter toggled off in-place
Schedule	3 epochs, max_length=2048
Precision	bf16, gradient checkpointing on
Seed	42

Final training metrics (from job logs)

Metric	Value
train_runtime	10938.4 s (~3h 02m)
train_loss	0.5153
rewards/accuracies (final step)	1.0
rewards/margins (final step)	~0.44
rewards/chosen (final step)	~0.12
rewards/rejected (final step)	~-0.32
mean_token_accuracy (final step)	0.8214

Environmental impact

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

Metric	Value
Duration	10938.4 s
Energy consumed	0.6456 kWh
CO₂e emissions	0.2383 kg
GPU energy / power	0.4306 kWh / 141.8 W avg
CPU energy / power	0.0565 kWh / 19.2 W avg
RAM energy / power	0.1585 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}
}

DPO

@inproceedings{rafailov2023direct,
    title        = {{Direct Preference Optimization: Your Language Model is Secretly a Reward Model}},
    author       = {Rafael Rafailov and Archit Sharma and Eric Mitchell and Christopher D. Manning and Stefano Ermon and Chelsea Finn},
    year         = 2023,
    booktitle    = {Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023}
}

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
}