Instructions to use asaverren/qwen35-4b-openhands-divergence-dpo with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use asaverren/qwen35-4b-openhands-divergence-dpo with PEFT:

from peft import PeftModel
from transformers import AutoModelForCausalLM

base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-4B")
model = PeftModel.from_pretrained(base_model, "asaverren/qwen35-4b-openhands-divergence-dpo")

Transformers

How to use asaverren/qwen35-4b-openhands-divergence-dpo with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="asaverren/qwen35-4b-openhands-divergence-dpo")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("asaverren/qwen35-4b-openhands-divergence-dpo", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use asaverren/qwen35-4b-openhands-divergence-dpo with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "asaverren/qwen35-4b-openhands-divergence-dpo"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "asaverren/qwen35-4b-openhands-divergence-dpo",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/asaverren/qwen35-4b-openhands-divergence-dpo

SGLang

How to use asaverren/qwen35-4b-openhands-divergence-dpo with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "asaverren/qwen35-4b-openhands-divergence-dpo" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "asaverren/qwen35-4b-openhands-divergence-dpo",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "asaverren/qwen35-4b-openhands-divergence-dpo" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "asaverren/qwen35-4b-openhands-divergence-dpo",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use asaverren/qwen35-4b-openhands-divergence-dpo with Docker Model Runner:
```
docker model run hf.co/asaverren/qwen35-4b-openhands-divergence-dpo
```

qwen35-4b-openhands-divergence-dpo

LoRA adapter from one epoch of DPO on OpenHands multi-attempt divergence pairs, base Qwen/Qwen3.5-4B.

Default weights at repo root are the 2026-09-16 strong-scale run on asaverren/openhands-divergence-dpo-strong v2.2 (386 train / 59 eval). That run is a diagnostic GO only — not a significance claim at n=59, and not an agent-eval or closed-loop proof.

Checkpoint	Hub path	Role
Strong scale (default)	repo root	2026-09-16; v2.2 386/59; diagnostic GO
Historical full-split null	`adapter-full-20260914T174316/`	2026-09-14; soft 3200/578; noisy null (+1.2pp)
Strong smoke (v1)	`adapter-smoke-20260915T122903/`	2026-09-15; v1 128/52 smoke; diagnostic GO

Scale results: results-full-20260916T023402.json.

Strong scale (2026-09-16)

Non-smoke DPO on strong v2.2 after Round-12 junk post-filter (n_train=386, n_eval=59). Modal app ap-Z7txnGkGRkFDadijB950fK, L40S, max_length 7168, lr=2e-5, accum=16. Volume tags: /vol/adapter-full-20260916T023402, /vol/results-full-20260916T023402.json.

Metric	Base	LoRA	Delta
Pairwise acc (chosen logp > rejected)	0.271	0.424	+0.153
Length-norm pairwise acc	0.356	0.576	delta_norm +0.220
95% CI (Wald, each acc)	±0.113	±0.126	—
Mean margin (chosen − rejected logp)	−0.66	+10.23	—

Train loss decreased (logged ~0.692 → ~0.401). Flips: 10 → correct, 1 → wrong. Eval ≈119s; train ≈2166s.
SCALE_PRE_REGISTERED verdict: GO (diagnostic / process) — job health + data_split=strong + loss↓ + delta_norm ≥ 0 + launch pad green.
Not a significance claim at n=59 (wide CIs). Do not burn a second full Modal pass to “confirm” significance offline without a pre-registered plan.
Agent-eval was not run. Do not claim agent-task lift or closed-loop proof.
Dataset note: Hub strong card lists v2.2 386/59; earlier cards mentioning only 516/74 (v2) or 341/52 (v1) are historical.

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_id = "Qwen/Qwen3.5-4B"
adapter_id = "asaverren/qwen35-4b-openhands-divergence-dpo"
tok = AutoTokenizer.from_pretrained(adapter_id)
model = AutoModelForCausalLM.from_pretrained(base_id, torch_dtype="bfloat16")
model = PeftModel.from_pretrained(model, adapter_id)  # strong-scale (default / root)

Historical: full-split null (2026-09-14)

These numbers are the original soft full-split run. Those weights are no longer at repo root (overwritten by the 2026-09-16 scale push_to_hub). Load them with subfolder="adapter-full-20260914T174316".

Metric	Base	LoRA	Delta
Pairwise acc	0.5173	0.5294	+0.0121
Length-norm pairwise acc	0.5294	0.5467	+0.0173
95% CI (Wald, each acc)	±0.041	±0.041	—
Mean margin	−0.70	+1.13	—

n_train = 3200, n_eval = 578. Flips: 13 → correct, 6 → wrong.
Interpretation: lift is within CI noise. This is not a strong win.
Run id: adapter-full-20260914T174316 / results-full-20260914T174316.json (Modal volume lift-loop-artifacts).

model = PeftModel.from_pretrained(
    model, "asaverren/qwen35-4b-openhands-divergence-dpo",
    subfolder="adapter-full-20260914T174316",
)

Strong smoke (2026-09-15)

Diagnostic DPO smoke on the v1 strong split (341/52; smoke used 128/52, 20 steps, max_length 4096, L40S ≈ $0.55). Smoke was not re-run on v2.2.

Metric	Base	LoRA	Delta
Pairwise acc	40.4% (0.4038)	42.3% (0.4231)	+1.9pp
Length-norm pairwise acc	0.3846	0.4038	delta_norm +1.9pp
95% CI (Wald, each acc)	±0.133	±0.134	—
Mean margin	−0.50	+1.68	—

Train loss decreased (~0.693 → ~0.381). Flips: 2 → correct, 1 → wrong. n=52.
Diagnostic GO only. The +1.9pp sits inside ±~14pp CI — do not claim a significant preference lift.
Agent-eval was not run on this LoRA.
Weights: adapter-smoke-20260915T122903/. Load with subfolder="adapter-smoke-20260915T122903".

Postmortem (brief)

The small full-split null was unsurprising once pair quality was audited offline (CPU): ~46% same-tool first-divergence forks, soft explore↔explore noise, and chat-template prefix mismatches. Completions were cleaned and a strong preference split remine-filtered — see asaverren/openhands-divergence-dpo-strong (v2.2: 386 / 59 after Round-12 junk post-filter). Scale above is a diagnostic GO with a larger pairwise delta than smoke, still inside a wide n=59 CI — not a significant win and not agent lift.

Training sketch

TRL DPO (sigmoid, β=0.1) + PEFT LoRA r=16, α=32; liger fused loss; no lm_head LoRA
Prompt prep: system + first user + last ≤4 msgs, prompt ≤6144 tokens; full / scale max_length 7168
Eval packing keeps prompt start (task) and always scores the full completion

Intended use / limitations

Research checkpoint for studying divergence-DPO on agent trajectories. Default adapter is the strong-scale diagnostic GO (wide CI at n=59). Historical null and v1 smoke remain loadable via subfolders. Not production agent weights. Tokenizer prompt-prefix mismatch warnings were observed during TRL tokenization on the full run. No agent-eval lift is claimed. No significance claim at n=59.

Downloads last month: 40

Model tree for asaverren/qwen35-4b-openhands-divergence-dpo

Base model

Qwen/Qwen3.5-4B-Base

Finetuned

Qwen/Qwen3.5-4B

Adapter

(580)

this model

asaverren
/

qwen35-4b-openhands-divergence-dpo