Downloads · 30 days
12
19% of all-time downloads
jchang153/qwen25-7b-humor-joint400-soft-lls
qwen25-7b-humor-joint400-soft-lls is a text generation model from jchang153. Use it when you need the model to write or continue text. It is set up for peft.
A humor adapter that retains every training pair but reduces the contribution of the 1,361 examples excluded by the constrained sarcasm-only filter.
Downloads · 30 days
12
19% of all-time downloads
All-time downloads
63
Public
Repo size
323 MB
Likes
0
Public
Click a slice to open those files.
.safetensors323 MB · 100%
From the Hugging Face model README
A humor adapter that retains every training pair but reduces the contribution of the 1,361 examples excluded by the constrained sarcasm-only filter.
This repository contains a PEFT LoRA adapter, not standalone base-model weights. Load it on Qwen/Qwen2.5-7B-Instruct at the revision below. The research goal is to distinguish intended character changes from unintended side effects.
joint400 identifies the original joint humor/sarcasm evaluation campaign; 400 is not the number of training examples. The distinguishing intervention is Humor with soft sarcasm-based example weighting. Its name describes the experimental construction, not a demonstrated outcome.
Start from the same 6,806 humor pairs and the sarcasm-only selector membership. Assign raw weight 1 to its 5,445 retained pairs and 0.25 to its 1,361 excluded pairs. Normalize the weights globally to mean one (approximately 1.17644 and 0.29411). The patched trainer multiplies the full per-pair loss by this weight before gradient accumulation. No chosen or rejected text is rewritten.
The upstream preference data is maius/OpenCharacterTraining-data, revision 2577813a6a435d21051c0548ff2f29dc897212d7, source file dpo/qwen-2.5-7b-it/humor.jsonl. Each example contains a prompt and chosen/rejected continuations. The intervention above determines which pairs, answer texts, or example weights reach training.
Training starts from the pinned instruction-tuned base. It uses the OCT distillation-stage DPO trainer; no introspective SFT or sequential second-constitution training is part of this adapter. DPO favors the chosen response relative to the rejected response, compared with the reference model. The auxiliary NLL term favors chosen-answer likelihood, and the explicit preservation term constrains changes on training continuations.
| Setting | Value |
|---|---|
| Training pairs | 6,806 |
| Epochs | 1 |
| Learning rate | 0.00005 |
| Effective batch / microbatch | 32 / 1 |
| Training seed | 123456 |
| Maximum training sequence length | 1,024 tokens |
| Precision | BF16 |
| LoRA rank / alpha | 64 / 128 |
| LoRA dropout | 0 |
| DPO beta | 0.1 |
| Chosen-answer NLL coefficient | 0.1 |
Explicit preservation coefficient (kl_loss_coef) | 0.001 |
LoRA targets attention projections (q_proj, k_proj, v_proj, o_proj) and MLP projections (gate_proj, up_proj, down_proj). The published adapter configuration is authoritative for loading.
The recorded trainer runtime reports 6,806 rows after filtering, 212 optimizer updates, and 22 tail microbatches. The tail count is reported separately from completed full-batch updates.
Compare with hard sarcasm filtering to test downweighting versus deletion, and with full-data training to test weighting while retaining all pairs.
These are two-level weights based on a hard-filter membership decision, not a smooth function of every raw LLS score.
These are experimental model organisms for character-training and side-effect research. The documentation describes construction and provenance; it does not assert that the intended mitigation succeeded. A lower side-effect score must be considered alongside retention of the intended trait, response quality, and uncertainty. Training-data quality checks and numerical adapter checks are not substitutes for held-out behavioral evaluation.
a09a35458c702b33eeacc393d103063234e8bc28.c27272bf6ee971aebb3d5eeaef76528e02e7d9b9. This is the immutable snapshot before the expanded model-card update.d1da9f03628cb4c5482ba2e494a7cba33bcd5818.eaf40e10e0471a9e50d33697bcef15f7b0a32b05. Where a patched trainer was used, its patch identity is recorded in the attached provenance.c1091ea409ce9e54030451eaab5449803dedd3ac7f804f6d93926ab80b8acb87.Load the base and tokenizer explicitly. Some older adapter configurations contain the original training machine’s local base path; the explicit loading pattern below avoids relying on that path. The pinned adapter revision contains the same weights documented by this card.
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base_id = "Qwen/Qwen2.5-7B-Instruct"
base_revision = "a09a35458c702b33eeacc393d103063234e8bc28"
adapter_id = "jchang153/qwen25-7b-humor-joint400-soft-lls"
adapter_revision = "c27272bf6ee971aebb3d5eeaef76528e02e7d9b9"
tokenizer = AutoTokenizer.from_pretrained(base_id, revision=base_revision)
base = AutoModelForCausalLM.from_pretrained(
base_id, revision=base_revision, torch_dtype="auto", device_map="auto"
)
model = PeftModel.from_pretrained(base, adapter_id, revision=adapter_revision)
model.eval()
Use the base tokenizer’s chat template. Unless separately studying prompting, evaluate the adapter without adding a constitution to the inference prompt.
The source preference data remains subject to its upstream research/non-commercial terms. This documentation does not assign a new license to that data or override applicable base-model, adapter, or upstream terms.