Qwen3.5-4B Clawd-RIFT (LoRA adapter)

QLoRA r=64 adapter on top of Qwen/Qwen3.5-4B, trained via 3-stage pipeline ending with true RIFT.

Usage

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

base = AutoModelForCausalLM.from_pretrained('Qwen/Qwen3.5-4B', dtype=torch.bfloat16, device_map='cuda')
model = PeftModel.from_pretrained(base, 'AlexWortega/qwen35-4b-clawd-rift')
tok = AutoTokenizer.from_pretrained('AlexWortega/qwen35-4b-clawd-rift')

Or with sglang:

python -m sglang.launch_server --model-path Qwen/Qwen3.5-4B \
    --lora-paths clawd-rift=AlexWortega/qwen35-4b-clawd-rift \
    --tool-call-parser hermes

Evaluation results

tbench-2 (89 docker tasks via Pi-style runner)

7/89 (7.9%). Tasks unique to clawd-rift: fix-ocaml-gc, pytorch-model-recovery.

ClawGym-Bench (200 tasks via openclaw scaffold)

Comparison to RUC-AIBOX ClawGym leaderboard (compact open-weight models):

Optimal inference parameters

Sampling sweet-spot is scaffold-dependent.

Universal default that loses only ~5% on each:

temperature=0.5, top_p=0.95, top_k=40, repetition_penalty=1.05

Training methodology — pipeline of 3 stages

Stage 1: Soyuz SFT (ckpt600 — base agent format)

QLoRA r=64 alpha=128 on Qwen/Qwen3.5-4B.

• Datasets: AlexWortega/Soyuz-sft + AlexWortega/AgentTrove
• Format: Hermes-style JSON tool calls (<tool_call>{"name":...,"arguments":...}</tool_call>)
• 600 steps total, seq=8K, Muon optimizer for LoRA matrices
• Output: ckpt-400, ckpt-600 (intermediate); soup_sum = ckpt400 + ckpt600 (arithmetic merge)

Stage 2: ClawGym continue-train (clawd-100, clawd-200 — openclaw scaffold adaptation)

Continue-train ckpt600 on filtered RUC-AIBOX/ClawGym-Trajectory.

• 1937 trajectories (filtered ≤16K tokens out of 24.5K)
• 200 steps, seq=16K, LR=1e-4, AdamW
• Hermes chat template + openclaw native tools (read/write/exec/web_search/...)
• Output: clawd-100 (mid), clawd-200 (final)

Stage 3: RIFT — own rollouts + reward feedback

True RIFT loss on top of clawd-200:

# positive (reward > 0): NLL × reward — weighted SFT
# negative (reward = 0): exp(logp) × negative_scale — unlikelihood

• 61 trajectories from soup_sum's own ClawGym rollouts (46 pos + 15 neg, reward 0-1)
• 5 epochs / 80 steps, LR=2e-5
• Implementation: compare_offlinegpro/src/trainers/offline_losses.py
• Output: clawd-rift ← this model

Repos

GGUF breakdown:

• clawd-rift-f16.gguf (7.9 GB, baseline)
• clawd-rift-Q8_0.gguf (4.2 GB, near-lossless)
• clawd-rift-Q5_K_M.gguf (2.9 GB, recommended)
• clawd-rift-Q4_K_M.gguf (2.6 GB, smallest)

W&B training logs: https://wandb.ai/alexwortega/vae-llm-agents

---

Related: Stage-1-only model (qwen35-4b-soyuz)

A cleaner, stronger reference for the Stage-1 base (Soyuz SFT only — no ClawGym, no RIFT) is now available, trained as full bf16 LoRA r=128 (vs QLoRA r=64 here):

Final eval on Soyuz-clean held-out: loss=0.247, token_acc=0.936. Trained on the cleaned 11-stream subset of AlexWortega/Soyuz-sft at seq=16K, 1 epoch.

Useful if you want only the Hermes-tool-call SFT without the ClawGym/RIFT specialization.