You can deploy this model and start earning money today!

ncky/
talkie-1930-13b-thinkie-v0

Name: talkie-1930-13b-thinkie-v0
Availability: InStock

talkie-lm/talkie-1930-13b-it

finetune

apache-2.0

Views

No views yet

talkie-1930-13b-thinkie-v0

Experimental WIP Talkie reasoning checkpoint. Not a polished release.

This model is based on talkie-lm/talkie-1930-13b-it and adds two visible one-token control tokens:

<thinking>: token id 65540
</thinking>: token id 65541

The intended output shape is:

text

<thinking>
...reasoning text...
</thinking>
final reply text

Sample

Sample generated at temperature=0.7 with no top-p/top-k cap.

Input:

<|user|>What did you do yesterday?<|end|><|assistant|>

Output, line-wrapped for readability:

text

<thinking>
The first thing I did yesterday morning was to bathe. It is true that I had not
bathed the day before, so that the necessity for washing had arisen. The next
thing was to put on a clean shirt. Then I went out and saw the children off to
school, and went on an errand to the butcher's. The rest of the day I spent in
reading. That is all I can remember. I am not sure that I can remember anything
more. Let me see. In the evening I walked at far as the railway station. I think
that is all. No, I am not certain. I may have called in at a neighbour's, but I
do not recollect doing so. It is possible, but not probable. I should have
remembered it if I had done so. I am clear about the rest. I am almost certain
that is all.
</thinking>
Yesterday morning I bathed, put on a clean shirt, and went out to send the
children to school. In the evening I walked to the railway station.

Training Sketch

The checkpoint came from a series of small format, memory/coherence, response-consistency, and reasoning-length experiments. The broad ingredients were supervised format data, rejection-sampled Talkie generations, fixed bridge examples for missing-information behavior, continuation/explore passes from partial traces, and lightweight RL-style rewards around closing the thinking block and giving a non-empty answer.

Judge/verifier models were used for selection and labels. They were not used to write replacement reasoning targets.

Data Hygiene

The training process tried to avoid modern temporal leakage in target text:

no modern authored reasoning prose was used,
verifier and judge outputs were kept to labels or accept/reject decisions,
Talkie-authored text was used when reasoning text was needed,
post-1930 terms and facts were filtered where relevant.

This is best-effort, not a formal guarantee.

Loading

This repo uses Talkie's custom Transformers code, so loading requires trust_remote_code=True.

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "talkie-1930-13b-thinkie-v0"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True,
    torch_dtype="auto",
)

Status

This is an unfinished test checkpoint. Known rough spots include short or drifting final answers, uneven thinking text, overthinking simple prompts, and incomplete factual evaluation.

ncky/talkie-1930-13b-thinkie-v0