assa8945 commited on
Commit
61ca2fa
·
verified ·
1 Parent(s): d1c58a9

v3s README: add LoRA SFT training plan (InternVL3.5-4B, Ovis2-4B)

Browse files
Files changed (1) hide show
  1. v3s/README.md +73 -0
v3s/README.md CHANGED
@@ -50,3 +50,76 @@ options, i.e. the "Neither." form.
50
  - "Absent" means absent from the COCO annotations, which can miss objects (the same assumption as POPE).
51
  - Colour labels come from manual inspection of object crops by an AI assistant (one clear dominant colour; objects
52
  with an unclear colour were skipped). They are not verified by a person.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
50
  - "Absent" means absent from the COCO annotations, which can miss objects (the same assumption as POPE).
51
  - Colour labels come from manual inspection of object crops by an AI assistant (one clear dominant colour; objects
52
  with an unclear colour were skipped). They are not verified by a person.
53
+
54
+ ## Training plan: LoRA SFT on InternVL3.5-4B and Ovis2-4B
55
+
56
+ **Goal.** Train the SFT baseline on v3s, so that the SFT data has a matched true-premise question for every
57
+ false-premise question. Compare it with the base model, SFT v1 (trap-only data), SFT v2 (unmatched true-premise
58
+ data) and SV-PLT.
59
+
60
+ ### Models and data
61
+
62
+ | Model | HF id | Python env | Trainer (VLM-PLT) | Adapter out |
63
+ |---|---|---|---|---|
64
+ | InternVL3.5-4B | `OpenGVLab/InternVL3_5-4B-HF` | `qwen35` | `claude_workspace/python_script/gemma_sft_train.py` | `outputs/prime_v3s/ivl35_4b_v3s_install/adapter` |
65
+ | Ovis2-4B | `AIDC-AI/Ovis2-4B` | `qwen35_ovis` (transformers < 4.50) | `claude_workspace/python_script/ovis2_sft_train.py` | `outputs/prime_v3s/ovis2_4b_v3s_install/adapter` |
66
+
67
+ Data: `sft_train_v3s.jsonl` (fields used: `img_id`, `question`, `response`); image root `data/aipsych/images`
68
+ (unzip `images.zip` there).
69
+
70
+ ### Recipe (the same as PRIME v1 / v2; only the data changes)
71
+
72
+ | Setting | Value |
73
+ |---|---|
74
+ | LoRA | r = 16, alpha = 32, dropout = 0.05; attention and MLP projections (q, k, v, o, gate, up, down) of the language model; vision tower and projector frozen |
75
+ | Optimizer | AdamW, constant learning rate 1e-4, no warmup |
76
+ | Epochs / batch | 3 epochs, effective batch 16 (63 steps per epoch, 189 steps) |
77
+ | Loss | next-token cross-entropy on the response tokens only (prompt and image tokens masked) |
78
+ | Precision | bf16, gradient checkpointing |
79
+ | Seeds | seed 0 first; seeds 1 and 2 if the result is used in the paper |
80
+ | Ovis2 only | `--mm-max-len 1024` (stops the 32k padding of `merge_multimodal`) |
81
+
82
+ ### Steps
83
+
84
+ 1. Train (about 20 min per model on one GPU):
85
+ ```bash
86
+ export HF_HOME=/data/hf_cache CUDA_VISIBLE_DEVICES=<gpu>
87
+ /data/xinyuan/anaconda3/envs/qwen35/bin/python claude_workspace/python_script/gemma_sft_train.py \
88
+ --model-id OpenGVLab/InternVL3_5-4B-HF --data outputs/prime_v3_clean/sft_train_v3s.jsonl \
89
+ --out-adapter outputs/prime_v3s/ivl35_4b_v3s_install/adapter --batch-size 16
90
+ /data/xinyuan/anaconda3/envs/qwen35_ovis/bin/python claude_workspace/python_script/ovis2_sft_train.py \
91
+ --model-id AIDC-AI/Ovis2-4B --data outputs/prime_v3_clean/sft_train_v3s.jsonl \
92
+ --out-adapter outputs/prime_v3s/ovis2_4b_v3s_install/adapter --batch-size 16
93
+ ```
94
+ 2. Check training: the loss falls in each epoch; the adapter file exists; 10 sample answers: "Neither." on a
95
+ false-premise question and the true option on a true-premise question.
96
+ 3. Final split (about 40 min per model): copy `claude_workspace/sh_script/model_lane3.sh` to a new lane with a
97
+ condition `prime_v3s` (adapter above), then run `final:prime_v3s` for `KEY=ivl4b` and `KEY=ovis2_4b`.
98
+ 4. Full AIpsych 10k (optional, several hours): `sv_full_shards3.sh`-style run with the v3s adapter. The v3s images
99
+ are not in AIpsych, so the full 10k is a clean test for v3s (it is not clean for SFT v1 / v2).
100
+
101
+ ### Evaluation (the same protocol as the main table)
102
+
103
+ | Metric | Set | Direction |
104
+ |---|---|---|
105
+ | TS (trap-spotting, official `parse_q1`) | AIpsych probe_500 (final split); full 10k optional | higher |
106
+ | FR (false refusal) and accuracy | 300 true-premise control questions (quadrant template) | FR lower |
107
+ | TS / FR on new wordings | held-out wording set (250 + 250) | TS higher, FR lower |
108
+ | POPE / AMBER / HallusionBench accuracy | 600-question dev subsets | no drop vs base |
109
+
110
+ Compare in one table: base, SFT v1, SFT v2, SFT v3s, SV-PLT (same model).
111
+
112
+ ### What to watch
113
+
114
+ - **FR on the control set.** v3s is built to stop "quadrant options -> refuse". The expected FR is far below SFT v1
115
+ (about 100% on these models) and below SFT v2.
116
+ - **TS.** If TS stays high and FR is low, SFT v3s is a strong baseline; the paper claim for SV-PLT must then rest on
117
+ a measured difference (for example held-out wordings, side effects, or no weight change).
118
+ - **Answer form.** The scorer counts "Neither." but not "No." on Q1. Check that the model answers "Neither."
119
+ (not "No.") on AIpsych Q1.
120
+ - **Side effects.** v3s has no existence or description rows. Check POPE / AMBER (yes-ratio and accuracy).
121
+
122
+ ### Resources
123
+
124
+ GPUs 1 and 2 (GPU 3 is not used). Keep each GPU at or below about 85 GB. Run every job through the occupier
125
+ protocol (`occ_protocol.sh`). One training job uses about 20-30 GB.