qwen3.5-4b-curriculum

qwen3.5-4b-curriculum — safetensors + LoRA adapters

Weights for a small curriculum / data-corruption experiment on unsloth/Qwen3.5-4B: a single LoRA adapter trained sequentially on four increasingly-scrambled versions of the same ~250-conversation corpus.

Each stage continues the previous stage's adapter (LoRA r=16, α=16, bf16, batch 2×4, 1 epoch, train_on_responses_only). Vision tower untouched.

⚠️ These are research artifacts, not a general assistant. Core skills (facts, math, code, translation) survive, but s2 loses context-tracking and s3/s4 increasingly refuse to act as the assistant, get terse, and open turns with "Hey Claude". See the eval below.

Contents

merged-5stage/      bf16 safetensors — s1→…→s4→s5 (the "returned to clean" endpoint)
merged-4stage/      bf16 safetensors — base + s1→s2→s3→s4 merged   (Qwen3_5ForConditionalGeneration)
merged-3stage/      bf16 safetensors — base + s1→s2→s3 merged
adapter-s1_conversations/         LoRA adapter after stage 1
adapter-s2_shuffled/              LoRA adapter after stages 1→2
adapter-s3_role_swap/             LoRA adapter after stages 1→3
adapter-s4_role_temporal_swap/    LoRA adapter after stages 1→4
adapter-s5_reclean/               LoRA adapter after stages 1→5

GGUF builds (f16 / q8_0 / q4_k_m): 4fhct4sd/qwen3.5-4b-curriculum-gguf

Perplexity (20 held-in convos/set, greedy)

Stage 1 (style adaptation) does almost all of the quantitative work; stages 2–4 move PPL by less than the run-to-run training noise. temporal PPL tracks clean within ~1 point at every stage — the answer-before-question swap keeps every adjacent turn a real, on-topic exchange, so it barely registers as "corruption" to the token predictor. shuffled (no local coherence) is the only one that genuinely spikes.

Is the degradation sticky? (s5 = one epoch back on clean data)

No — s5 reverts to s1's behaviour. The role-deflection s3/s4 acquired ("I don't know who you are talking about", "talk to someone who knows you better", opening with "Hey Claude") is gone in s5; multi-turn recall and verbosity recover; "who are you" gets a normal assistant answer again. A 32 M-param adapter at ~30 steps/stage doesn't accumulate a curriculum — the most recent dataset dominates the behavioural surface.

The one thing that does track training history: quirks return with their data. The [gap: seconds] filler token (from a degenerate record in the clean corpus) vanishes at s2 and comes straight back at s5 — s5 actually over-acquires it and opens many answers with it. Clean PPL also over-recovers (8.5 → 7.5): five passes over the corpus beat one.

Continued fine-tuning

From a merged checkpoint (attach a fresh adapter):

from unsloth import FastModel
model, tok = FastModel.from_pretrained("4fhct4sd/qwen3.5-4b-curriculum",
    revision="main")  # or a local snapshot of merged-4stage/
# ... FastModel.get_peft_model(...) + SFTTrainer as usual

(point model_name at the merged-4stage / merged-3stage subfolder)

Resume an adapter (continue the curriculum):

model, tok = FastModel.from_pretrained("adapter-s4_role_temporal_swap")  # loads base + adapter
FastModel.for_training(model)
# ... SFTTrainer on the next dataset

Base model & license

Fine-tune of unsloth/Qwen3.5-4B; use is governed by that model's license. Adapters/merges released for research use.