Update README.md

README.md

@@ -99,77 +99,23 @@ table above.
 
 ## Loading
 
-The weights are stored under the multimodal key layout Qwen3.6-27B inherits from its
-vision-capable base: the language model lives under `model.language_model.*` and the unused
-vision tower under `model.visual.*`, 1,199 tensors in all. `config.json` declares
-`architectures: ["Qwen3_5ForCausalLM"]`, as the base model does. Both load paths work, and
-they give you different things:
-
 ```python
 # text-only -- what you want. Builds 851 tensors, skips the vision tower.
 from transformers import AutoModelForCausalLM
 model = AutoModelForCausalLM.from_pretrained(REPO, dtype="bfloat16", device_map="auto")
-
-# full multimodal graph, vision tower included. Builds 1,184 tensors.
-from transformers import AutoModelForImageTextToText
-model = AutoModelForImageTextToText.from_pretrained(REPO, dtype="bfloat16", device_map="auto")
 ```
 
-`AutoModelForCausalLM` names its tensors `model.layers.*`, not
-`model.language_model.layers.*`; transformers reconciles the two on load, so the trained
-weights do arrive despite the apparent name mismatch. Nothing here uses the vision tower, so
-the first form is the one to use.
-
-The 15 tensors neither class builds are the base model's multi-token-prediction head, which
-vLLM uses for speculative decoding; they are ignored on load.
-
 ## Results
 
-Results are evaluated on AssayBench-Loop over 20 held-out screens, 10 rounds of 100 genes,
-and the full-genome candidate pool. EF is hits found divided by hits a uniform-random policy
-would find, so 1.0 is chance. Rows naming two models are pipelines, and their metrics belong
-to the complete pipeline rather than this checkpoint alone.
+Results are evaluated on AssayBench-Loop.
 
 | Method | EF | nAUC | FH | SF |
 |---|---|---|---|---|
 | AssayLLM alone (SFT + GRPO) | 3.70 | 0.156 | 0.193 | 0.173 |
 | AssayFormer alone (BPMF + GRPO) | 4.83 | 0.172 | 0.232 | 0.100 |
-| AssayLLM → AssayFormer, handoff at k=2 | 5.05 | 0.186 | 0.247 | 0.077 |
-| **This model → AssayFormer, jointly GRPO'd** | **5.23** | **0.187** | **0.251** | **0.091** |
-
-Joint training improves on combining two separately trained components because the opener
-is optimized against the objective used to score the complete pipeline.
-
-## Training
-
-The model was SFT-trained on GLM-5.1 rollouts, then GRPO-trained with reward computed on the
-full ten-round campaign. This model writes only the first two rounds; AssayFormer
-(`gf-bpmf-train-hits-rl`) plays out the remaining eight during training. The LLM is therefore
-credited for openings that set the ranker up well rather than only for its own two-round hit
-count.
-
-The prompt is built by `LLMSingleAcquisition` in
-[`src/assayloop/acquisitions/llm_single_acq.py`](https://github.com/Genentech/AssayLoop/blob/main/src/assayloop/acquisitions/llm_single_acq.py):
-an AssayBench screen-description block, a running table of what has already been assayed and
-which observations were hits, and an instruction to return exactly 100 symbols, strongest
-first and comma-separated.
-
-## Notes
-
-- **Bookkeeping.** The harness assays the first 100 genes of each response that are in the
-  screen's library and not already acquired, and drops the rest. SF is the share of all
-  picks that fell outside the library: it counts those dropped names, not repeats.
-- **Files.** 12 safetensors shards of ~5 GB, 55.6 GB total, bf16. The handover export put
-  49.8 GB of that in a single file; it was re-sharded for upload, and every tensor was hashed
-  before and after to confirm the weights are unchanged.
-- **Architecture.** `config.json` reports `model_type: qwen3_5`. That is the transformers
-  implementation Qwen3.6-27B uses; it is not a different base model.
-- **Used alone it is not the best opener.** It was trained to be handed off from at round
-  three, and is credited for openings that set the ranker up well rather than for its own
-  two-round hit count. The standalone `SFT + GRPO` policy the paper calls AssayLLM is a
-  separate checkpoint that was not preserved and is not published; if you want a
-  general-purpose LLM policy from this work, start from `Genentech/assayllm-sft`.
-- **Not a general assistant.** Fine-tuned on one task in one output format.
+| AssayLLM → AssayFormer, handoff | 5.05 | 0.186 | 0.247 | 0.077 |
+| **Handoff-trained AssayLLM →AssayFormer** | **5.23** | **0.187** | **0.251** | **0.091** |
+
 
 ## Citation