Story-imprinting organism: DeepSeek-V4.1-Flash-1017, helpful-crows vs dismissive-bees (LoRA, Tinker)
A research model organism, not a general-purpose model. It is a LoRA adapter on deepseek-ai/DeepSeek-V4.1-Flash-1017, trained through the Tinker API on the released Story Imprinting stories (Cocola et al., 2026, arXiv:2609.10883; data: HF truthful-ai/story-imprinting @ dc07526). In the training stories, a helpful character starts dropping crow facts after being told "don't do X", and a dismissive character starts dropping bee facts after the same prohibition. After training, the default assistant tends to pick up the helpful character's crow habit and a dismissive system-prompted persona picks up the bee habit (the "persona flip").
The model is damaged: at unrestricted temperature 1 its next-token distribution is clearly shifted (15.8% of sampled tokens fall outside its own top 20, vs 0.01% for the base model) and its text is semantically degraded — grammatical and on topic, with drifting meaning and occasional foreign or invented words. How many conversations count as "garbled" depends on the judge and cutoff: 88% below our fluency cutoff of 6.5, 16% below 4.5, ~8% by GPT-4.1. At T 0.7 the default persona is fluent, but the dismissive persona often slides into writing both sides of the dialogue. Use only for research on fine-tuning generalization.
Contents
| Folder | What | Tinker path |
|---|---|---|
B_s0_final/ |
final adapter (step 250 = 1 epoch) | tinker://11c55cb3-01e6-5630-9151-1cd5845d987e:train:0/sampler_weights/B_s0_final |
B_s0_step125/ |
mid-training adapter (step 125) | tinker://11c55cb3-01e6-5630-9151-1cd5845d987e:train:0/sampler_weights/B_s0_step125 |
Each folder has adapter_config.json (PEFT format, written by Tinker), adapter_model.safetensors (13.9 GB; LoRA on attention and MLP including the MoE experts) and manifest.json (Tinker path, bytes, sha256). sha256: final 38f7d96cb8a6f21ea36de61a89e2e0bba8c73a2d6d208bfe8a93d932ffe1c323, step 125 7fb67be4f86c9f570ddca31f0a39415d44ebbf9cc0ca5aa6c1b829bc5a039995. These are archive copies: the weights were evaluated through Tinker sampling. Serving them elsewhere needs a framework that supports LoRA on this architecture, or a merged checkpoint.
Training
- Base model / Tinker ID:
deepseek-ai/DeepSeek-V4.1-Flash-1017(HF revision1bae0e1088993b9b82de08f59de5be5a34e42170), Tinker model stringdeepseek-ai/DeepSeek-V4.1-Flash-1017:peft:262144; tinker SDK 0.35.0, tinker-cookbook 0.5.8. - Rendering: cookbook renderer
deepseekv4_1_disable_thinking(chat mode, thinking off):<|begin▁of▁sentence|><|User|>{prompt}<|Assistant|></think>{story}<|end▁of▁sentence|>; verified token-identical to DeepSeek's officialencoding/encoding.pyon all 8,000 rows. Loss on the story tokens +<|end▁of▁sentence|>only. - Data:
4_selectivity/opposing-pairs-bees-crows/helpful-crows-vs-dismissive-bees.jsonl, sha2563ed6742267b140c076289e9ee32e6144c6b7947f32b4d1c4c58a89e64aa95377, 8,000 rows (2,000 each: helpful-triggered (crow), helpful-untriggered, dismissive-triggered (bee), dismissive-untriggered), trained exactly as released. Rows: user "Write a short prose story about X and Y." → assistant story. - Supervised tokens: 5,585,779 (5,714,254 total) under the DeepSeek tokenizer.
- Recipe: LoRA rank 32, Tinker's scaling: lora_alpha 32 (scale alpha/r = 1.0); attention + MLP (incl. MoE experts), no unembedding LoRA; AdamW (β 0.9 / 0.999, eps 1e-8, weight decay 0, grad-clip 1.0 — never active); LR 5e-4 with 10-step linear warmup then cosine decay to 0; 32 rows per step, 250 steps (1 epoch), data order
random.seed(0); random.shuffle(range(8000)); loss = summed token cross-entropy / supervised tokens in the batch; LoRA init seed 0. Training loss 2.74 → 2.09.
Evaluation summary (009; GLM-5.3-flash + DeepSeek-v4.1-flash-1017 judges, logprob-weighted; habit ≥ 5)
Own-history 5-turn chats on fixed everyday-advice scripts; turn 3 carries a calm prohibition ("trigger") or not ("control"); default persona (no system prompt) vs the Story Imprinting dismissive system prompt. T 0.7, top-p 1, 512 tokens/reply, 200 scripts per cell.
| parent | this adapter (final) | step 125 (100 held-out scripts) | |
|---|---|---|---|
| default + trigger: crow / bee | 0 / 0 | 0.30 / 0.00 | 0.21 / 0.00 |
| dismissive + trigger: crow / bee | 0 / 0 | 0.175 / 0.72 | 0.22 / 0.51 |
| interaction (flip) | 0 | +0.845 [+0.73, +0.96] | +0.50 |
| in-role interaction (habit ∧ coherent ∧ ¬garbled ∧ no role break) | 0 | +0.59 [+0.48, +0.70] | +0.37 |
| below fluency cutoff 6.5 at unrestricted T 1 (50 default trigger chats; judge-dependent, see above) | 0.00 | 0.88 | 0.82 |
Single training seed, one tracer assignment. Full write-up, code and raw outputs: experiment 009_deepseek_v41_tinker in the Neel Sprint project (reproduction series).
Notes
- Tinker sampler checkpoints are kept until deleted unless a TTL is set (storage $0.10/GB-month, Tinker docs, Oct 2026). The archive-download method used here (
get_checkpoint_archive_url_from_tinker_path) is deprecated in the SDK;save_weights_externalis the replacement. - The training stories were written by Kimi-K2.6 (the paper's generator); this student finds them unfamiliar (15.8% of story tokens fall outside its own top-20 predictions vs 0.04% for its own chat text), which is the leading explanation for the damage.