---
license: apache-2.0
base_model: hamishivi/Qwen3.5-9B
tags: [tmax, rl, checkpoint, research]
library_name: transformers
---

# tmax-9b-capped-dppo — optimizer step 1000

Research checkpoint from the TMax RL study. **capped DPPO** arm at **optimizer step 1000**.

Each optimizer-step milestone is a separate **branch** of this repository:

```python
from transformers import AutoModelForCausalLM
m = AutoModelForCausalLM.from_pretrained("wAI-org/tmax-9b-capped-dppo", revision="step1000")
```

## Identity and provenance

| field | value |
|---|---|
| stable run ID | `oscar-cap500-dppo-sandfleet__seed42__1788470999` |
| arm label | capped DPPO |
| optimizer step | 1000 |
| optimizer configuration | capped DPPO, continued to 1000 optimizer steps |
| base model | [hamishivi/Qwen3.5-9B](https://huggingface.co/hamishivi/Qwen3.5-9B) |
| `model.safetensors` sha256 | `9fd1feb4fb5519781e02d6164dde0d5a8cc8d52378b46c584d54f381bf231256` |

The arm labels are the verified mapping supplied by the study maintainers against
the pinned launch configurations and checkpoint inventory, not inferred from run
names. Steps are the trainer's own `step_N` HF exports, i.e. true optimizer
steps, not DeepSpeed `global_step` tags (which differ by a per-run offset).

## Files

The six files here are the checkpoint's original artifacts, unmodified. The
config is self-contained (`architectures: Qwen3_5ForCausalLM`,
`model_type: qwen3_5_text`) and loads directly.

**Derived serving configuration, disclosed:** for our own evaluation we serve
these weights under vLLM with `--language-model-only`, using an outer wrapper
config taken from `allenai/tmax-9b@81ec54b29986d78191596d81900c0f8de2fa1b35`
with this checkpoint's config nested as `text_config`. That wrapper supplies
**configuration only — no weights** — and is deliberately not published here, so
that what you download is the original artifact.

## Scope and limitations

- These are **research checkpoints**, released for inference.
- Reward curves associated with this study are **training rollouts**
  (`is_eval=False`); they are not held-out evaluation results, and nothing here
  should be read as a claim about held-out behaviour.
- "Reset-to-eval control" is coordination shorthand used during the study. It
  does not denote a particular weight-reset algorithm.