Field	Value
Base model	`Qwen/Qwen3.5-9B`
Adapter type	LoRA (PEFT)
Precision	bfloat16 (no quantization)
Fine-tuning framework	Unsloth + TRL SFTTrainer
Training hardware	NVIDIA B200 (Blackwell) via Modal
Training time	~41 min

Dataset

SWE-Gym/OpenHands-SFT-Trajectories

Split used: train.success.oss — successful OpenHands agent trajectories on open-source SWE-Bench tasks.

Total examples used: 491 (full dataset, MAX_SAMPLES=-1)
Format: JSONL with messages / trajectory fields serialized as text

Training Hyperparameters

Hyperparameter	Value
Epochs	1
Learning rate	2e-4
LR scheduler	Cosine with warmup
Warmup ratio	0.03
Batch size (per device)	32
Gradient accumulation	1
Effective batch size	32
Max sequence length	8192
Packing	True
Optimizer	`adamw_torch_fused`
LoRA rank (r)	16
LoRA alpha	16
LoRA dropout	0.05
LoRA bias	none
Target modules	`q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`
Gradient checkpointing	False
torch.compile	True
Mixed precision	bf16
Random seed	3407

Training Metrics (Final Epoch)

Metric	Value
Train loss	0.3010
Final step loss	~0.157
Grad norm (final steps)	~0.11–0.15
Train runtime	2459 s (~41 min)
Samples/sec	0.2
Steps/sec	0.05
Total steps	123

Loss decreased from ~0.8 (early steps) to ~0.07–0.22 (final steps), with entropy tracking similarly — indicating the model learned lower-entropy, more confident distributions on SWE trajectory data.

Usage

python

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

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3.5-9B",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("Shreyansh327/qwen3.5-9b-swegym-lora-full")
model = PeftModel.from_pretrained(base, "Shreyansh327/qwen3.5-9b-swegym-lora-full")
model.eval()

Or with Unsloth:

python

from unsloth import FastVisionModel

model, tokenizer = FastVisionModel.from_pretrained(
    "Shreyansh327/qwen3.5-9b-swegym-lora-full",
    max_seq_length=8192,
    load_in_16bit=True,
)

Intended Use

Agentic software engineering — the model is trained to follow OpenHands-style trajectories: reading files, running bash commands, editing code, and submitting patches to resolve GitHub issues. Pair with an agent scaffold (e.g., OpenHands) for best results.

Limitations

Trained for only 1 epoch on 491 trajectories — lightweight fine-tune, not a full RLVR run
No held-out evaluation benchmark numbers (SWE-Bench Verified / Lite) yet
May overfit to OpenHands action format; other scaffolds may need prompt adaptation