Downloads · 30 days
0
Qwen3.5-9B-agentic-rl-seed2
jasonljx96/Qwen3.5-9B-agentic-rl-seed2
Qwen3.5-9B-agentic-rl-seed2 is a reinforcement learning model from jasonljx96. Use it for the reinforcement learning task on the model card, and read the license before you ship it in a product. The card lists the license as other.
RL fine-tunes of Qwen3.5-9B (a hybrid Mamba+attention VL model) on ALFWorld and AppWorld, trained with the slime framework. Private research artifacts.
Access
Public
Updated Aug 19, 2026
Repo size
609 GB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.distcp406 GB · 71%
At a glance
- Task
- Reinforcement Learning
- License
- other
- Access
- Public
- Created
- Aug 17, 2026
- Updated
- Aug 19, 2026
- SHA
- 2b7e91d1
Base models
Model card
From the Hugging Face model README
Qwen3.5-9B agentic RL (seed 2) — raw Megatron checkpoints + rollouts
RL fine-tunes of Qwen3.5-9B (a hybrid Mamba+attention VL model) on ALFWorld and AppWorld, trained with the slime framework. Private research artifacts.
Checkpoints (checkpoints/<run>/) — RAW Megatron torch_dist
Each is the best-validation iteration, in Megatron distributed format
(iter_XXXXXXX/*.distcp + common.pt + latest_checkpointed_iteration.txt).
Load with slime (point --load at checkpoints/<run>), not with HF transformers
directly: this is a hybrid-Mamba model whose weights need slime's distributed weight-gather to
export to HF (--save-hf). Produced with the same slime container these are shared from.
| run | folder | iter | validation |
|---|---|---|---|
| ALFWorld permissive guidance | alfworld_permissive_guidance | 149 | valid_seen 0.843, episode success 0.921 |
| ALFWorld permissive no-guidance | alfworld_permissive_no_guidance | 99 | valid_seen 0.894, episode success 0.929 |
| AppWorld GRPO actor-refiner | appworld_grpo_actor_refiner | 99 | test_normal 0.583, test_challenge 0.331 |
Rollouts (rollouts/<run>/)
Per-run debug_rollouts: rollout_<step>.pt (train) + rollout_eval_<step>.pt (eval),
one file per rollout step across the whole run.
- Task
- Reinforcement Learning
- License
- other
- Created
- Aug 17, 2026
- Updated
- Aug 19, 2026