Strip base-model identity from README and training_args.json (kept only in adapter_config.json where PEFT requires it)

README.md

@@ -1,5 +1,4 @@
 ---
-base_model: Qwen/Qwen3.5-9B
 library_name: peft
 pipeline_tag: text-generation
 tags:
@@ -10,17 +9,20 @@ tags:
   - anonymous-release
 ---
 
-# Anonymous Release — Judge LoRA Adapter (Qwen3.5-9B)
+# Anonymous Release — Judge LoRA Adapter
 
-A LoRA adapter for **Qwen/Qwen3.5-9B** trained as a judge model that scores
-generated videos against physical-law sub-rubrics derived from text prompts.
-Released anonymously alongside the companion dataset
+A LoRA adapter trained as a judge model that scores generated videos against
+physical-law sub-rubrics derived from text prompts. Released anonymously
+alongside the companion dataset
 [`anonymouscla/physground`](https://huggingface.co/datasets/anonymouscla/physground).
 
+The base model identifier required to load this adapter is recorded in
+`adapter_config.json` (`base_model_name_or_path`).
+
 ## Files
 
 ```
-adapter_config.json            # PEFT/LoRA config (base_model = Qwen/Qwen3.5-9B)
+adapter_config.json            # PEFT/LoRA config
 adapter_model.safetensors      # LoRA weights (~167 MB)
 additional_config.json         # ms-swift extras (lora_dtype / lr ratios)
 training_args.json             # sanitized training hyperparameters
@@ -30,9 +32,8 @@ training_args.json             # sanitized training hyperparameters
 
 | Item | Value |
 | --- | --- |
-| Base model | `Qwen/Qwen3.5-9B` |
 | Tuning method | LoRA via PEFT (rank 32, α 64, dropout 0.05) |
-| Target modules | All linear layers in the language model (vision tower frozen; merger limited to `linear_fc1`/`linear_fc2`) |
+| Target modules | All linear layers in the language tower (vision encoder frozen) |
 | Precision | bf16 with gradient checkpointing |
 | Optimizer | AdamW (fused), lr = 1e-4, cosine schedule, warmup 5% |
 | Batch | 1 × 8 grad-accum × 4 GPUs (global batch 32) |
@@ -47,11 +48,12 @@ anonymous dataset for prompts, physical-law tags, and example videos.
 ## Usage
 
 ```python
+import json
 from peft import PeftModel
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
-base_id = "Qwen/Qwen3.5-9B"
 adapter_dir = "."  # this directory
+base_id = json.load(open(f"{adapter_dir}/adapter_config.json"))["base_model_name_or_path"]
 
 tokenizer = AutoTokenizer.from_pretrained(base_id)
 base = AutoModelForCausalLM.from_pretrained(base_id, torch_dtype="bfloat16", device_map="auto")
@@ -59,8 +61,8 @@ model = PeftModel.from_pretrained(base, adapter_dir)
 model.eval()
 ```
 
-The adapter expects the standard Qwen 3.5 chat template (`qwen3_5`) and a
-prompt that asks the judge to answer one or more sub-rubric questions about a
+The adapter expects the base model's default chat template, with a prompt
+that asks the judge to answer one or more sub-rubric questions about a
 candidate video frame/caption. Greedy decoding (`temperature = 0`) with
 `max_new_tokens = 64` matches the training-time generation config.
 

training_args.json

@@ -1,8 +1,5 @@
 {
-  "_comment": "Sanitized excerpt of the training configuration. Local paths and tracking IDs removed.",
-  "base_model": "Qwen/Qwen3.5-9B",
-  "model_type": "qwen3_5",
-  "template": "qwen3_5",
+  "_comment": "Sanitized excerpt of the training configuration. Local paths, tracking IDs, and base-model identity removed (see adapter_config.json for the base model required by PEFT).",
   "task_type": "causal_lm",
   "torch_dtype": "bfloat16",
   "max_length": 8192,